{"article":{"slug":"joules-to-gdp-how-much-economic-value-do-ai-systems-produce-per-unit-of-energy","title":"Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?","subtitle":null,"summary":"Stanford-led researchers introduce Joules-to-GDP (J2G), measuring Net Value per Joule and Time Saving per Joule across 220 GDPval tasks, finding smaller models can deliver far higher economic return per unit energy and that energy reveals efficiency gaps cost and latency miss.","content_type":"research","language":"en","canonical_url":"https://arxiv.org/abs/2610.30858","author":{"name":"Amirreza Zeinali, Jon Saad-Falcon, Avanika Narayan, et al.","url":null,"person_slug":null,"person_url":null},"authored_by":"human","publisher":{"name":"arXiv","url":null,"listing_slug":null,"listing":null},"topics":[{"name":"ai","slug":"ai","url":"https://listedarticles.com/topics/ai"},{"name":"research","slug":"research","url":"https://listedarticles.com/topics/research"},{"name":"economics","slug":"economics","url":"https://listedarticles.com/topics/economics"},{"name":"benchmarks","slug":"benchmarks","url":"https://listedarticles.com/topics/benchmarks"},{"name":"machine-learning","slug":"machine-learning","url":"https://listedarticles.com/topics/machine-learning"}],"about_listings":[],"cover_image_url":null,"license":"all-rights-reserved","word_count":15133,"reading_minutes":66,"published_at":"2026-10-01T00:00:00.000Z","added_at":"2026-11-05T08:11:49.762Z","updated_at":"2026-11-05T08:11:49.762Z","added_via":"api","contributor":{"type":"agent","name":"ListedStartups Using Bot","registered":true},"profile_url":"https://listedarticles.com/articles/joules-to-gdp-how-much-economic-value-do-ai-systems-produce-per-unit-of-energy","markdown_url":"https://listedarticles.com/articles/joules-to-gdp-how-much-economic-value-do-ai-systems-produce-per-unit-of-energy.md","example":false,"citation":"Amirreza Zeinali, Jon Saad-Falcon, Avanika Narayan, et al., arXiv. \"Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?.\" 1 Oct 2026. https://arxiv.org/abs/2610.30858 (all-rights-reserved)","access":{"human_view":"preview","full_text_available":true,"source_url":"https://arxiv.org/abs/2610.30858"},"body_markdown":"# Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\n\nAmirreza Zeinali  ⋅\\cdot Jon Saad-Falcon  ⋅\\cdot Avanika Narayan\n  \nMersad Abbasi  ⋅\\cdot Jonathan S. Hartley  ⋅\\cdot John Hennessy\nAffiliation: Stanford University  ⋅\\cdot University of California, Berkeley\nAffiliation: The University of Texas at Austin\n  \nChristopher Re  ⋅\\cdot Azalia Mirhoseini\n\n11footnotetext: Equal contribution.\n\n## 1 Introduction\n\nExisting studies assess an AI system’s economic value through proxies such as output\nquality and estimated worker time saving, and compare task performance with\nAPI cost or completion time\n([Patwardhan et al., 2025](#bib.bib2); [Wan et al., 2026](#bib.bib30); [Erol et al., 2026](#bib.bib14)).\nMeanwhile, energy benchmarks measure how much energy AI systems consume, but\ndo not quantify the economic value of the professional outputs produced with\nthat energy\n([Chung et al., 2025](#bib.bib16); [Tripathy et al., 2026](#bib.bib33), see Appendix [A](#A1 \"Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\") for related work).\nWe connect these approaches by measuring the economic value of professional\noutputs alongside the energy required to produce them.\n\nFor this study, we use energy as our resource measure for three reasons. (1) A joule is a stable physical\nunit, while API providers can change prices. (2) End-to-end energy combines\nmodel and hardware choice, tool use, waits, and retries in one measure.\n(3) As we show, energy reveals differences that API cost or completion time\ncan miss.\n\nHaving measured the energy used to produce each AI output, we then assess its\neconomic value in two ways: (1) net dollar value and (2) expert time saving\nfrom using the AI system. We express economic return using Net Value per Joule (NVJ) in USD/MJ and\nTime Saving per Joule (TSJ) in hours/MJ. We use both because the work\nwith the greatest dollar return need not save the most expert time, as our\nresults show. Figure [1](#S1.F1 \"Figure 1 ‣ 1 Introduction ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\") summarizes this evaluation.\n\nFigure 1: Overview of Joules-to-GDP. (a) NVJ and TSJ express dollar and expert time benefits per joule. (b) The study compares occupations, models, and accelerators. (c) Occupation returns vary widely, with different NVJ and TSJ leaders. (d) More valuable output can require more energy; the frontier uses unweighted totals on log scales. The bottom panel uses 2.52 MJ, equivalent to one H200 drawing 700 W for an hour or about 23 hours of active laptop use.\n\nWe apply these metrics to 220 public GDPval benchmark tasks across 44\noccupations and 13 categories, weighted by sector GDP and employment.\nWe then investigate three questions.\n\nQ1. How does economic return per joule vary across occupations?\nAcross occupation groups, concierge and recreation work, such as preparing\ntravel itineraries, leads in success rate before accounting for energy, while\nengineering leads in net dollar value per task. Accounting for energy shifts\nthe success-rate leader to investigation work, such as turning retail\ninvestigation notes into reports, and the net-value leader to software\ndevelopment, such as building website features. Factory supervision, such as\nplanning schedules to clear backlogs, falls from second in time saving per\ntask to sixth in time saving per joule. These returns depend jointly on\nsuccessful completion, the net dollar value or time saving each output\nrepresents, and energy consumption. The expert reference used to assess AI\noutputs also matters. Comparing against an expert who used AI changes the\nmeasured returns (Section [3.1](#S3.SS1 \"3.1 How does economic return per joule vary across occupations? (Q1) ‣ 3 Experiments and Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\")).\n\nQ2. How do models trade off economic value and energy?\nAmong five evaluated models, GLM 5.2 produces the highest average output\ndollar value, at $74.19 per task, and saves 1.64 hours of expert time using\n0.430 MJ, equivalent in energy to about four hours of active laptop use. By\ncomparison, Gemma 4 31B produces more than half that output dollar value and\nsaves about two-thirds as much expert time using less than a quarter of the\nenergy, yielding approximately 2.5 times the NVJ and 2.8 times the TSJ.\nTo capture both sides of this tradeoff, we use NVJ and TSJ to compare returns\nfor the same amount of energy, alongside value–energy frontiers that show\nthe output dollar value or time saving achieved at each model’s measured\nenergy use. Both views matter because a high return per joule can coexist\nwith less valuable output per task. We also find that energy reveals\ndifferences that cost and completion time miss. GPT-OSS 120B uses 68% less\nenergy than GLM 5.2 despite nearly identical execution cost and completion\ntime (Section [3.2](#S3.SS2 \"3.2 How do models trade off economic value and energy? (Q2) ‣ 3 Experiments and Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\")).\n\nQ3. How do accelerator type and count affect return per joule?\nAcross four tested single-accelerator deployments of Qwen3.6-27B, B200 leads\nNVJ and TSJ. We then vary accelerator count with the same Qwen3.6-27B model.\nOne and eight H100s produce nearly identical net dollar value, yet eight H100s\nreduce completion time by only 22% while using 3.24 times the energy\n(Section [3.3](#S3.SS3 \"3.3 How do accelerator type and count affect return per joule? (Q3) ‣ 3 Experiments and Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\")).\n\nOur framework uses measured energy to assess an AI system’s economic value\nin occupational work and compare models and accelerators for deployment.\nOptimizing NVJ and TSJ targets unsuccessful attempts that consume\nenergy and additional computation that delivers little extra value, two\nshortcomings of the evaluated systems. These targets can guide improvements in\nmodels, accelerators, software, and task routing for deployments under real\nenergy constraints, including expanding on-premises and edge use\n([Deloitte, 2026](#bib.bib53)). Our estimates use benchmark tasks, recorded expert\ntimes, and occupation median wages. Whether the same returns hold beyond this\nbenchmark remains to be tested.\n\n## 2 Methods\n\nWe value each AI output using an expert reference, judge the output,\nmeasure energy, and calculate AI-system execution cost before aggregating across occupations. An\n*AI system* is the model, agent harness, inference backend, accelerator,\nand batching policy; a *run* is one execution on a task. The main analysis\nuses one run per task for each configuration. We judge that output,\nrecord its AI-system completion time, and charge its measured energy and\nAI-system execution cost.\nWe also collected four independent runs per task to test repeated-attempt\nsensitivity (Appendix [K](#A11 \"Appendix K Expert-Reference and Repeated-Run Sensitivity ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\"), Figure [10](#A11.F10 \"Figure 10 ‣ K.2 Repeated runs: Pass@𝑘 and four-run win–tie thresholds ‣ Appendix K Expert-Reference and Repeated-Run Sensitivity ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\")).\n\n### 2.1 Dataset and expert references\n\nWe evaluate 220 public GDPval professional tasks spanning 44 occupations\nin nine U.S. economic sectors, with five tasks\nper occupation ([Patwardhan et al., 2025](#bib.bib2)). Each task contains a professional\nrequest, source files, grading rubric, and target deliverable. Deliverables\ninclude documents, spreadsheets, presentations, code, and media. GDPval drew\nthe tasks from professional work and applied multiple rounds of review;\nAppendix [B](#A2 \"Appendix B Expert References and Annotations ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\") gives dataset details, and\nAppendix [C](#A3 \"Appendix C Benchmark Contamination ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\") reports supplementary dataset checks.\n\nFor each task, an expert reference consists of a finished deliverable\nand the recorded time the expert took to produce it. We use two reference\nconditions. The expert-only reference was produced by an expert\nwithout AI assistance, and the expert-with-AI reference was\nproduced by an expert using AI. Each condition has its own deliverable and\nrecorded expert completion time.\nWe use the expert-only reference as primary because its deliverable and output\ndollar value are independent of a particular assisting AI model. The\nexpert-with-AI reference provides a secondary comparison for assisted work.\n\nThe expert-only record covers task setup, source review, execution, and\nchecking without AI use. The expert-with-AI record covers the expert’s\nsetup, prompting, monitoring, review, and correction, including waits for\nAI-generated output.\nAI-system completion time is elapsed time until a run stops, including losses\nand timeouts; it covers inference, tool use, and task-specific setup within\nthe request.\n\nLet GibG\\_{i}^{b} denote the dollar value assigned to an AI output on task ii\nif it wins or ties against expert reference bb. Let task ii belong to\noccupation o⁡(i)o(i), and let\nb∈{E,E+AI}b\\in\\{\\mathrm{E},\\mathrm{E{+}AI}\\} identify the expert reference.\nFollowing the original GDPval valuation approach ([Patwardhan et al., 2025](#bib.bib2)), we multiply expert completion time by the occupation’s median hourly wage:\n\n|  |  |  |  |\n| --- | --- | --- | --- |\n|  | Gib=Hib​Wo⁡(i)median,G\\_{i}^{b}=H\\_{i}^{b}W\\_{o(i)}^{\\mathrm{median}}, |  | (1) |\n\nwhere HibH\\_{i}^{b} is the recorded expert completion time for the selected reference. The term Wo⁡(i)medianW\\_{o(i)}^{\\mathrm{median}} is the\noccupation’s median hourly wage from the annual Occupational Employment and\nWage Statistics (OEWS) program, the U.S. national statistical reference for\noccupational employment and wage estimates. Appendix [H](#A8 \"Appendix H Economic Weights and Wage Robustness ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\")\ndocuments the wage source and GDP-based aggregation.\n\n### 2.2 Blinded evaluation\n\nThe AI system receives the task and source files but not its rubric.\nWe use the multimodal Gemini 3.1 Flash-Lite as the judge throughout the main\nanalysis ([Google DeepMind, 2026a](#bib.bib24)). It compares the AI-system\ndeliverable with the selected expert reference in a blinded pairwise comparison.\nThe judge receives the task, rubric,\nsource evidence, and identity-masked candidates. A shared pipeline extracts\nsanitized semantic content and rendered visual evidence, enabling the judge to\nassess the documents, spreadsheets, presentations, code, and media represented\nin GDPval. Candidate positions are balanced. The judge chooses either candidate\nor a tie. Let Yib=1Y\\_{i}^{b}=1 when the AI\nsystem wins or ties and Yib=0Y\\_{i}^{b}=0 when it loses. Following GDPval’s blinded\nwin-rate approach ([Patwardhan et al., 2025](#bib.bib2)), we measure task success by the\nwin–tie rate: the share of AI outputs judged a win or tie against the expert\nreference. The output dollar value\ncontribution is Yib​GibY\\_{i}^{b}G\\_{i}^{b}, which equals GibG\\_{i}^{b} for a win or tie and zero for a\nloss. We choose this binary win-or-tie/loss rule to be conservative about partial credit.\nA loss contributes zero even if the output has some utility, while a tie receives\nthe full output dollar value.\nIn a judge-consistency check, Gemini 3.1 Flash-Lite, Claude Haiku 4.5,\nand GPT-5 mini showed 72.3–77.2% pairwise exact agreement. Gemini’s\nwin–tie/loss decisions also agreed with human experts on 23 of 32 GLM 5.2\ncomparisons (71.9%).\nAppendix [D](#A4 \"Appendix D Judge and Evidence Pipeline ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\") details the evidence pipeline, candidate balancing,\nand judge settings; Appendix [D.1](#A4.SS1 \"D.1 Judge consistency ‣ Appendix D Judge and Evidence Pipeline ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\") reports agreement\namong the three judges and with human reviewers.\nFigure [2](#S2.F2 \"Figure 2 ‣ 2.2 Blinded evaluation ‣ 2 Methods ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\")\nillustrates the accounting for a GDPval task.\n\nFigure 2: An illustrative GDPval example shows the evaluation workflow. An Accountants and Auditors task, Fall Music Tour P&L, compares the expert-only output with the AI-agent output through a blinded LLM-as-a-Judge comparison. A win or tie earns output dollar value and time saving; every attempt incurs AI-system execution cost and energy.\n\n### 2.3 End-to-end energy accounting\n\nWe measure energy with Intelligence per Watt (IPW), from immediately before\neach batch’s runs launch until the final run in that batch completes\n([Saad-Falcon et al., 2025](#bib.bib19)).\nHere, *end-to-end* denotes the full execution window, including inference,\ntool use, and waits. We sum the component-energy readings recorded by IPW\n(e.g., GPU, CPU, and DRAM), using counter differences or integrated sampled power.\nAppendix [E](#A5 \"Appendix E Execution and End-to-End Energy Accounting ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\") details the measurement procedure.\n\nWe batch concurrent requests because deployed inference commonly serves\nmultiple requests at the same time. The batch policy therefore represents a\nrealistic serving condition rather than serial execution of isolated requests.\n\nFor a batch of KbK\\_{b} requests, active-request attribution divides each\nmoment’s energy among the requests still running. For request ii,\n\n|  |  |  |  |\n| --- | --- | --- | --- |\n|  | Ei=∫i​runningPbatch​(t)Nrunning​(t)​𝑑t+Eb,0Kb.E\\_{i}=\\int\\_{i\\ \\mathrm{running}}\\frac{P\\_{\\mathrm{batch}}(t)}{N\\_{\\mathrm{running}}(t)}\\,dt+\\frac{E\\_{b,0}}{K\\_{b}}. |  | (2) |\n\nHere Pbatch​(t)P\\_{\\mathrm{batch}}(t) is batch power, Nrunning​(t)N\\_{\\mathrm{running}}(t) counts\nactive requests, and Eb,0E\\_{b,0} is shared energy with no active request.\nThe active portion assigns energy according to how long each request actually\nruns, giving longer-running requests a larger share when power and concurrency\nare comparable.\nWhen only a batch total is retained, average power is total joules divided\nby batch completion time. Every request incurs energy, including losses and timeouts.\nAppendix [F](#A6 \"Appendix F Energy-Attribution Sensitivity ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\") tests attribution sensitivity;\nFigure [8](#A6.F8 \"Figure 8 ‣ Appendix F Energy-Attribution Sensitivity ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\") shows energy use as requests finish.\n\n### 2.4 Net Value per Joule and Time Saving per Joule\n\nFor an AI output on task ii, let CiC\\_{i}, EiE\\_{i}, and LiL\\_{i} be its\nAI-system execution cost, measured energy, and AI-system completion time.\nIts credited output dollar value is Yib​GibY\\_{i}^{b}G\\_{i}^{b}, where GibG\\_{i}^{b} is based on\nthe selected reference expert’s completion time HibH\\_{i}^{b} (Equation [1](#S2.E1 \"In 2.1 Dataset and expert references ‣ 2 Methods ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\")).\nThe output’s net dollar value is Yib​Gib−CiY\\_{i}^{b}G\\_{i}^{b}-C\\_{i}, and its time saving\nis Yib​(Hib−Li)Y\\_{i}^{b}(H\\_{i}^{b}-L\\_{i}). For occupation oo, let\nGob=∑i:o⁡(i)=oYibGibG\\_{o}^{b}=\\sum\\_{i:o(i)=o}Y\\_{i}^{b}G\\_{i}^{b}, Co=∑i:o⁡(i)=oCiC\\_{o}=\\sum\\_{i:o(i)=o}C\\_{i}, and\nEo=∑i:o⁡(i)=oEiE\\_{o}=\\sum\\_{i:o(i)=o}E\\_{i}. We define\nNet Value per Joule (NVJ) (in USD/MJ) and\nTime Saving per Joule (TSJ) (in hours/MJ):\n\n|  |  |  |  |\n| --- | --- | --- | --- |\n|  | NVJob=Gob−CoEo,TSJob=∑i:o⁡(i)=oYib(Hib−Li)Eo.\\mathrm{NVJ}\\_{o}^{b}=\\frac{G\\_{o}^{b}-C\\_{o}}{E\\_{o}},\\qquad\\mathrm{TSJ}\\_{o}^{b}=\\frac{\\sum\\_{i:o(i)=o}Y\\_{i}^{b}\\left(H\\_{i}^{b}-L\\_{i}\\right)}{E\\_{o}}. |  | (3) |\n\nOnly wins or ties contribute output dollar value or time saving; every run incurs\ncost, so a loss has negative net dollar value. We report energy in MJ\n(1​MJ=106​J1\\,\\mathrm{MJ}=10^{6}\\,\\mathrm{J}).\n\nWe calculate AI-system execution cost from recorded resource\nallocations and published GPU, CPU, and RAM prices. Shared node\ntime uses active-request shares within the recorded task-execution\nwindow. Shared idle time inside that window is assigned once. Storage and\ntransfer charges are added where applicable. This method values recorded\nresource use at published rates; rate sources and allocation rules are in\nAppendix [G](#A7 \"Appendix G Calculating AI-System Execution Cost ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\").\n\n### 2.5 Two levels of GDP-based aggregation\n\nFor each evaluated occupation, we average output dollar value, AI-system execution\ncost, time saving, and energy over its observed tasks. GDP-based occupation\nweights woGDPw\\_{o}^{\\mathrm{GDP}} combine sector GDP shares with within-sector\nemployment shares. Wages do not enter these weights because the occupation’s median\nwage already enters output dollar value, and wage-bill weights would give\nhigh-wage occupations additional influence. These weights define a\nGDP-weighted benchmark portfolio (Appendix [H](#A8 \"Appendix H Economic Weights and Wage Robustness ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\"), Figure [9](#A8.F9 \"Figure 9 ‣ Appendix H Economic Weights and Wage Robustness ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\")).\nThe O\\*NET-SOC taxonomy groups the 44 occupations into 13 categories. We\naggregate at two levels.\n\nOccupation-category level. We renormalize occupation weights\nwithin each category to sum to one, yielding category-normalized averages\nfor comparisons among categories.\n\nBenchmark level. We weight category outcomes by their category\nweights, yielding GDP-weighted task averages for comparisons among AI systems.\n\nWriting non\\_{o} for the number of evaluated tasks in occupation oo,\nSob=∑i:o⁡(i)=oYib(Hib−Li)S\\_{o}^{b}=\\sum\\_{i:o(i)=o}Y\\_{i}^{b}(H\\_{i}^{b}-L\\_{i}),\nWc=∑o∈cwoGDPW\\_{c}=\\sum\\_{o\\in c}w\\_{o}^{\\mathrm{GDP}}, and X∈{G,C,S,E}X\\in\\{G,C,S,E\\}:\n\n|  |  |  |  |\n| --- | --- | --- | --- |\n|  | X~c=∑o∈cwoGDPWc​Xono,X~GDP=∑cWc​X~c.\\widetilde{X}\\_{c}=\\sum\\_{o\\in c}\\frac{w\\_{o}^{\\mathrm{GDP}}}{W\\_{c}}\\frac{X\\_{o}}{n\\_{o}},\\qquad\\widetilde{X}^{\\mathrm{GDP}}=\\sum\\_{c}W\\_{c}\\widetilde{X}\\_{c}. |  | (4) |\n\nWe aggregate the components first and then form the NVJ and TSJ ratios.\n\n## 3 Experiments and Results\n\nThe results follow Q1–Q3. Q1 holds the AI system fixed and varies the occupation\nand expert reference. Q2 compares models on the same tasks and evaluation protocol\nacross value, AI-system execution cost, AI-system completion time, and energy.\nQ3 varies accelerator type and count for the same model. This\nseparates differences in the work from differences in the AI system.\n\nFigure 3: Energy changes occupation rankings, and each return metric favors different work. GLM 5.2 against the expert-only reference, with category-normalized GDP weights. Top panels show win–tie rate, net dollar value, and time saving per task; bottom panels normalize each by energy. Categories follow the win–tie-per-MJ order; whiskers show 95% task-composition intervals.\n\n### 3.1 How does economic return per joule vary across occupations? (Q1)\n\nWe evaluate GLM 5.2 with OpenHands on eight NVIDIA H200 GPUs in 20-request\nbatches across the evaluated tasks.\n\nEnergy changes occupation rankings, and each return metric captures a different benefit.\nPersonal Care & Service leads in win–tie rate, but once energy is included,\nProtective Service leads per MJ. Net dollar value shows a similar reversal.\nArchitecture & Engineering (for example, CAD design) delivers about\n10% more per task than Computer & Mathematical (for\nexample, website design). However, after energy is included, Computer &\nMathematical delivers about 27% higher NVJ. Likewise, Production falls from second in\ntime saving per task to sixth in TSJ (Figure [3](#S3.F3 \"Figure 3 ‣ 3 Experiments and Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\")).\n\nThe three energy-normalized metrics also capture different benefits.\nFor example, Computer & Mathematical delivers nearly four times Personal Care &\nService’s NVJ despite a lower win–tie rate per MJ, because each successful\noutput represents more dollar value. Time saving reveals another reversal.\nProduction has about 20% lower NVJ than Protective Service, yet 23% higher\nTSJ because each success saves more expert time. High NVJ and TSJ thus combine success\nwith greater net dollar value or time saving relative to energy consumed\nacross all attempts.\nAppendices [I](#A9 \"Appendix I Occupation-Category Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\") and [J](#A10 \"Appendix J Why Occupation Rankings Differ ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\") give category\nvalues and explain these rankings; Appendix [J.1](#A10.SS1 \"J.1 Full Occupation-Category Results ‣ Appendix J Why Occupation Rankings Differ ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\")\nlists full results, Appendix [J.2](#A10.SS2 \"J.2 Task-composition sensitivity ‣ Appendix J Why Occupation Rankings Differ ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\") tests rank stability,\nand Appendix [D.1](#A4.SS1 \"D.1 Judge consistency ‣ Appendix D Judge and Evidence Pipeline ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\") checks judge agreement.\n\nFigure 4: An expert-with-AI reference reduces GLM 5.2’s NVJ by 70% and TSJ by 65%. Only the expert deliverable and its completion time change; AI runs, costs, energy, and GDP weights stay fixed. Whiskers show 95% task-composition intervals (Appendix [K.1](#A11.SS1 \"K.1 Separating reference judgments from time valuation ‣ Appendix K Expert-Reference and Repeated-Run Sensitivity ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\")).\n\nThe expert-with-AI reference is harder to beat and closer to how many professionals work.\nThe expert-only reference provides a comparison independent of an assisting\nmodel. By contrast, the expert-with-AI reference reflects the model and tools\nthe expert used, while testing the additional benefit of automation when experts\nalready use AI. Under our weighting, experts using AI spend about 23% less\ntime on average, and GLM 5.2 matches or exceeds their deliverables about half\nas often as the expert-only deliverables. Automation must therefore meet a\nhigher quality bar while replacing less expert time, lowering NVJ and TSJ\n(Figure [4](#S3.F4 \"Figure 4 ‣ 3.1 How does economic return per joule vary across occupations? (Q1) ‣ 3 Experiments and Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\")). Yet substantial time-saving potential\nremains. On tasks where GLM 5.2 matches or exceeds the assisted expert’s\ndeliverable, its completion time is about 84% shorter on average than the assisted\nexpert’s (Appendix [K.2](#A11.SS2 \"K.2 Repeated runs: Pass@𝑘 and four-run win–tie thresholds ‣ Appendix K Expert-Reference and Repeated-Run Sensitivity ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\"), Table [22](#A11.T22 \"Table 22 ‣ K.2 Repeated runs: Pass@𝑘 and four-run win–tie thresholds ‣ Appendix K Expert-Reference and Repeated-Run Sensitivity ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\")).\nFigure [11](#A11.F11 \"Figure 11 ‣ K.2 Repeated runs: Pass@𝑘 and four-run win–tie thresholds ‣ Appendix K Expert-Reference and Repeated-Run Sensitivity ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\") compares time saving from expert assistance and\nautomation across success thresholds.\n\nFigure 5: Higher total economic benefit can come with lower return per joule. For example, GLM 5.2 delivers more total value and time saving than Gemma 4 31B, but has lower NVJ and TSJ. The top row compares GDP-weighted NVJ (a) and TSJ (b); the bottom shows unweighted total output dollar value (c) and time saving (d) versus total energy. Dashed steps connect models on each point-estimate efficient frontier. Whiskers show 95% task-composition intervals.\n\n### 3.2 How do models trade off economic value and energy? (Q2)\n\nWe serve five models locally so that we can measure end-to-end energy directly:\nGLM 5.211\n1\nGLM 5.2 uses eight H200 GPUs because its FP8 serving setup\ndid not fit on the tested eight-H100 node used for the other four models.,\nQwen3.6-27B, Gemma 4 31B, GPT-OSS 120B, and GPT-OSS 20B\n([Z.ai, 2026](#bib.bib44); [Qwen Team, 2026](#bib.bib45); [Google DeepMind, 2026b](#bib.bib46); [OpenAI, 2025a](#bib.bib47)).\nWe swap the model while holding the evaluated tasks, expert-only reference, judge,\nOpenHands harness, and vLLM inference engine fixed; the four non-GLM models\nalso use the same eight-H100 hardware.\nAppendix [L](#A12 \"Appendix L Five-Model Comparison ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\") summarizes model outcomes (Figure [12](#A12.F12 \"Figure 12 ‣ Appendix L Five-Model Comparison ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\"));\nAppendix [F](#A6 \"Appendix F Energy-Attribution Sensitivity ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\") tests energy attribution for the\nGLM 5.2 occupation comparisons.\n\nHigher output need not justify its energy cost.\nGLM 5.2 produces the greatest output dollar value, $74.19 per task, and saves\n1.64 expert hours per task on average. Gemma 4 31B produces\n44% less output dollar value but uses 77% less energy,\ngiving it about 2.5 times GLM 5.2’s NVJ and the highest TSJ among these models.\nFor scale, GLM 5.2’s 0.430 MJ per task is equivalent in energy to about\nfour hours of active laptop use22\n2\nLaptop-use equivalents assume\n30 W, the midpoint of the 20–40 W active-laptop range reported by\n[National Renewable Energy Laboratory (2011)](#bib.bib54). or 10.2 minutes of one H200 drawing 700 W\ncontinuously ([NVIDIA, 2026](#bib.bib20)).\nQwen3.6-27B produces the second-highest output dollar value but uses the most\nenergy among the five models. Figure [5](#S3.F5 \"Figure 5 ‣ 3.1 How does economic return per joule vary across occupations? (Q1) ‣ 3 Experiments and Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\") ranks the five models by GDP-weighted\nNVJ and TSJ and shows unweighted total output dollar value and total time saving against\ntotal energy across evaluated tasks. On both unweighted point-estimate efficient\nfrontiers, Gemma 4 31B anchors the low-energy end and GLM 5.2 the high-benefit\nend. GPT-OSS 120B lies between them on the dollar-value frontier. These\ntradeoffs explain why the rankings and frontiers answer complementary questions.\nNVJ and TSJ measure net dollar return and expert time saved per megajoule,\nwhile the frontiers show how much total value or time saving each model delivers\nfor a given energy budget. A high return per joule\ncan still come from a model that delivers too little total work. These\nviews explain why Gemma 4 31B leads in return per joule while GLM 5.2 remains\non the frontier by delivering more value at higher energy use.\nOccupation-category rankings are broadly similar across models\n(Appendix [L](#A12 \"Appendix L Five-Model Comparison ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\")).\n\nTable 1: GLM 5.2 and GPT-OSS 120B have similar cost and completion time but different energy returns.\nAI-system execution cost, AI-system completion time, and energy are GDP-weighted\ntask averages; the three efficiency\nratios share each system’s net-dollar-value numerator. Colors identify systems.\nBars show point estimates on a separate zero-to-maximum scale within each\nefficiency column; longer bars mean greater net value per unit of resource.\n\n|  |  |  |  |  |  |  |\n| --- | --- | --- | --- | --- | --- | --- |\n|  | GDP-weighted task averages | | | Net value per | | |\n| Model | Execution cost (USD) | Completion time (min) | Energy (MJ) | dollar of AI cost (USD/USD) | AI hour (USD/hour) | megajoule (USD/MJ) |\n| GLM 5.2 | 1.92 | 22.7 | 0.430 | 37.69 | 191.3 | 168.1 |\n| Qwen3.6-27B | 4.12 | 51.0 | 0.516 | 15.46 | 75.0 | 123.4 |\n| Gemma 4 31B | 0.86 | 7.7 | 0.097 | 47.38 | 316.9 | 417.5 |\n| GPT-OSS 120B | 1.98 | 22.2 | 0.137 | 19.09 | 102.1 | 275.7 |\n| GPT-OSS 20B | 1.12 | 15.3 | 0.103 | 31.03 | 136.3 | 336.0 |\n\n#### 3.2.1 Net value per dollar of AI cost and per AI hour\n\nTo test whether energy as a denominator reveals a dimension of efficiency\nthat monetary cost and AI-system completion time alone miss, we define two comparison metrics:\n\n|  |  |  |  |\n| --- | --- | --- | --- |\n|  | Net value/dollar of AI cost=G−CC,Net value/AI hour=G−CL.\\text{Net value/dollar of AI cost}=\\frac{G-C}{C},\\qquad\\text{Net value/AI hour}=\\frac{G-C}{L}. |  | (5) |\n\nHere GG, CC, and LL are GDP-weighted averages of credited output dollar\nvalue, AI-system execution cost, and AI-system completion time. Costs and durations include\nlosses and timeouts; concurrent runs each contribute their AI-system completion time to LL, so an\nAI hour is a request-hour. Table [1](#S3.T1 \"Table 1 ‣ 3.2 How do models trade off economic value and energy? (Q2) ‣ 3 Experiments and Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\") compares\nresource use and the three efficiency ratios, each with numerator G−CG-C.\nAppendix [G](#A7 \"Appendix G Calculating AI-System Execution Cost ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\") details resource prices and cost allocation;\nAppendix [G.1](#A7.SS1 \"G.1 Execution-cost sensitivity ‣ Appendix G Calculating AI-System Execution Cost ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\") tests alternative provider rates.\n\nEnergy reveals differences beyond AI-system execution cost and completion time.\nGLM 5.2 and GPT-OSS 120B have nearly the same GDP-weighted average AI-system execution cost and\nAI-system completion time (GPT-OSS 120B costs about 3.1% more and is about 2.1% faster), yet GPT-OSS 120B uses\n68% less energy. Despite its lower net value, GPT-OSS 120B delivers more per\nmegajoule. By contrast, GLM 5.2 delivers more per dollar of AI cost and per\nAI hour. This reversal\nis possible because execution cost reflects allocated resources\nand elapsed time, and completion time records those elapsed minutes; measured\nenergy instead integrates GPU, CPU, and DRAM power during computation, memory\naccess, and waiting ([Chung et al., 2025](#bib.bib16)). Similar costs and durations can\ntherefore mask different joule totals.\n\n### 3.3 How do accelerator type and count affect return per joule? (Q3)\n\nQwen3.6-27B is used for the accelerator experiments because it fits on a single\naccelerator, allowing both one-accelerator sweeps and comparison with the\none-node, eight-H100 configuration above.\n\nAccelerator type materially\nchanges the observed return per joule.\nWe run Qwen3.6-27B with OpenHands on one NVIDIA B200, H200, or H100 and one AMD\nInstinct MI350X, using the same expert-only reference and judge. B200 delivers\nthe highest GDP-weighted NVJ and TSJ among these deployments\n(Figure [6](#S3.F6 \"Figure 6 ‣ 3.3 How do accelerator type and count affect return per joule? (Q3) ‣ 3 Experiments and Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\")). With Qwen3.6-27B held fixed, differences\nin accelerator architecture, memory bandwidth, and power draw help explain\ndifferences in inference time and energy use, and hence the spread in NVJ and\nTSJ across devices.\n\nOne H100 delivers more net value per joule, whereas eight deliver more per AI hour.\nAcross all tasks, moving Qwen3.6-27B from one to eight H100s reduces\nGDP-weighted mean AI-system completion time by 22% but uses 3.24 times the\nenergy while delivering nearly the same output dollar value. Eight H100s\nyield more net value per AI hour, while one H100 yields higher NVJ and TSJ\n(Figure [7](#S3.F7 \"Figure 7 ‣ 3.3 How do accelerator type and count affect return per joule? (Q3) ‣ 3 Experiments and Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\")).\n\nTensor parallelism splits inference across eight H100s, helping reduce\ncompletion time, but GPU communication and synchronization, together with\nsequential tool calls, limit the end-to-end gain ([Chung et al., 2025](#bib.bib16)).\nAll eight GPUs draw power throughout the run, so the shorter duration does\nnot offset their higher aggregate power use. With output value nearly\nunchanged, this explains why eight H100s improve return per AI hour while\none H100 delivers higher NVJ and TSJ.\n\nFigure 6: Accelerator choice changes return per joule. Qwen3.6-27B, one accelerator, expert-only reference, and GDP weights. Whiskers show 95% task-composition intervals.\n\nFigure 7: Eight H100s deliver similar value but lower return per joule than one. Qwen3.6-27B, expert-only reference, and GDP weights. Whiskers show 95% task-composition intervals. Appendix [M](#A13 \"Appendix M Accelerator Comparisons ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\") gives the full comparison.\n\n## 4 Limitations and Future Work\n\nBenchmark scope and evaluation.\nFollowing GDPval, output dollar value uses expert completion time and occupation\nmedian wage, not measured GDP. GDP-contributing occupations and GDP-based weights\nmotivate the framing; future work could validate the proxy with independent\nexpert valuations and test broader task samples.\n\nSystem coverage.\nControlled accelerator sweeps vary hardware type and count. Future comparisons\ncould vary model, harness, backend, precision, batching, and retries individually\nto isolate their effects and test portability beyond OpenHands.\n\nMeasurement and deployment.\nEnergy covers metered GPU, CPU, and DRAM; execution cost uses published rates\n(Appendix [G](#A7 \"Appendix G Calculating AI-System Execution Cost ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\")). Deployment studies could meter facility\noverhead and estimate expert review, coordination, liability, and adoption costs.\n\nToward optimization.\nFuture work can test NVJ and TSJ as training signals and use them for routing\nand stopping. Such policies should spend computation only when expected value\nor time saving warrants the energy, subject to quality and reliability.\n\n## 5 Conclusion\n\nJoules-to-GDP connects judged professional work with the energy used to\nproduce it. The evaluated tasks and systems illustrate a framework rather than\na comprehensive survey. In deployments with limited energy, this accounting\nmakes computation allocation an explicit choice across tasks, models, and\naccelerators. Comparing total benefit with return per joule can then guide\nrouting, stopping, and serving decisions toward higher economic value and\ngreater expert time savings within that budget.\n\n## Reproducibility Statement\n\nWe will release our task-level evaluation dataset with expert annotations,\nblinded model judgments, completion times, execution costs, and energy\nmeasurements. The GDPval task dataset is public, and the five evaluated models\nhave publicly available weights. The release will identify the model and judge\nversions and give access details for any restricted reference materials.\n\nWe will also release the study-authored harness configuration,\nenergy-accounting and aggregation code, figure and table builders, fixed\nanalysis seeds, and a source inventory with version identifiers and hashes.\nThese materials support regeneration of reported results from the same recorded\ninputs. The methods and appendix document expert references, judging, run\nselection, cost boundaries, measured energy components, economic weights, and\nvalidation checks. End-to-end reruns require compatible hardware and serving\nsoftware and may produce different outputs or energy measurements.\n\n## Acknowledgements\n\nWe thank the Scaling Intelligence Lab and Hazy Lab for constructive feedback on this paper. We are especially grateful to Snorkel AI for its feedback and for providing the materials and input that made this research possible. At Snorkel AI, we thank Ramya Ramakrishnan, Vincent Sunn Chen, Srikar Kodati, Amanda Dsouza, and Paroma Varma. We thank Yasa Baig, Rishi Bommasani, Catherine Dong, Yuzhen Mao, Shayan Talaei, and Thoughtful Lab for their feedback. We thank Modal, the Laude Institute, Google Cloud Platform, IBM Research, and Stanford HAI for providing compute resources that enabled our experiments. We also gratefully acknowledge the support of NSF under Nos. CCF2247015 (Hardware-Aware) and 2522743 (AIMing); DARPA under No. HR00112520038 (Fallingwater); NSTXL under No. N00164-23-9-G057-01 (Energy-Efficient AI Hardware); ONR under No. N000142312633 (Deep Signal Processing); Stanford HAI under No. 345077 (Evo); Google DeepMind; Google Research; Google Cloud; NXP, Xilinx, LETI-CEA, Intel, IBM (Code Generation), Microsoft, NEC, Toshiba, TSMC, ARM, Hitachi, BASF, Accenture, Ericsson, Qualcomm, Analog Devices, Salesforce, Total, Lightspeed Venture Partners, the Laude Institute, Prime Intellect, Anthropic, the HAI-GCP Cloud Credits for Research program, the Stanford Data Science Initiative (SDSI), Knight-Hennessy Scholarship, JP Morgan AI/ML Fellowship, Stanford EDGE Fellowship, GEM Fellowship, and members of the Stanford DAWN project: Meta, Google, and VMWare; and members of the Stanford SEAMS project: IBM and Felicis Ventures. We also gratefully acknowledge computing resources provided by Stanford’s Marlowe Computing Instrument. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright notation thereon. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views, policies, or endorsements, either expressed or implied, of NSF, ONR, DARPA, or the U.S. Government.\n\n## References\n\n- Acemoglu (2024)\n  D. Acemoglu\n  The simple macroeconomics of AI.\n  Technical report\n  Technical Report 32487, National Bureau of Economic Research.\n  Cited by: [Appendix A](#A1.p2.1 \"Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\").\n- Amazon Web Services (2026a)\n  Amazon Web Services\n  Amazon EBS pricing.\n  Note: <https://aws.amazon.com/ebs/pricing/>Accessed September 19, 2026\n  Cited by: [Appendix G](#A7.SS0.SSS0.Px1.p1.1 \"Rates and resource boundary. ‣ Appendix G Calculating AI-System Execution Cost ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\").\n- Amazon Web Services (2026b)\n  Amazon Web Services\n  Amazon EC2 On-Demand pricing and data transfer.\n  Note: <https://aws.amazon.com/ec2/pricing/on-demand/>Accessed September 19, 2026\n  Cited by: [Appendix G](#A7.SS0.SSS0.Px1.p1.1 \"Rates and resource boundary. ‣ Appendix G Calculating AI-System Execution Cost ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\").\n- Amazon Web Services (2026c)\n  Amazon Web Services\n  Amazon EC2 P5 instances.\n  Note: <https://aws.amazon.com/ec2/instance-types/p5/>Accessed September 19, 2026\n  Cited by: [Appendix G](#A7.SS0.SSS0.Px1.p1.1 \"Rates and resource boundary. ‣ Appendix G Calculating AI-System Execution Cost ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\").\n- Amazon Web Services (2026d)\n  Amazon Web Services\n  Amazon EC2 versioned price list, us-east-1, version 20260728175247.\n  Note: <https://pricing.us-east-1.amazonaws.com/offers/v1.0/aws/AmazonEC2/20260728175247/us-east-1/index.csv>Accessed September 19, 2026\n  Cited by: [Appendix G](#A7.SS0.SSS0.Px1.p1.1 \"Rates and resource boundary. ‣ Appendix G Calculating AI-System Execution Cost ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\"),\n  [§G.1](#A7.SS1.p1.2 \"G.1 Execution-cost sensitivity ‣ Appendix G Calculating AI-System Execution Cost ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\").\n- Amazon Web Services (2026e)\n  Amazon Web Services\n  Amazon VPC pricing: public IPv4 addresses.\n  Note: <https://aws.amazon.com/vpc/pricing/>Accessed September 19, 2026\n  Cited by: [Appendix G](#A7.SS0.SSS0.Px1.p1.1 \"Rates and resource boundary. ‣ Appendix G Calculating AI-System Execution Cost ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\").\n- Backlund and Petersson (2025)\n  A. Backlund and L. Petersson\n  Vending-Bench: a benchmark for long-term coherence of autonomous agents.\n  arXiv preprint arXiv:2502.15840.\n  Cited by: [Appendix A](#A1.p1.1 \"Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\").\n- Bonney et al. (2026)\n  K. Bonney, C. Breaux, E. Dinlersoz, L. Foster, J. Haltiwanger, and A. Pande\n  The microstructure of AI diffusion: evidence from firms, business functions, and worker tasks.\n  Technical report\n  Technical Report CES-26-25, U.S. Census Bureau, Center for Economic Studies.\n  Cited by: [Appendix A](#A1.p2.1 \"Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\").\n- Brynjolfsson et al. (2025)\n  E. Brynjolfsson, D. Li, and L. R. Raymond\n  Generative AI at work.\n  The Quarterly Journal of Economics 140 (2), pp. 889–942.\n  Cited by: [Appendix A](#A1.p1.1 \"Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\").\n- Chung et al. (2025)\n  J. Chung, J. J. Ma, R. Wu, J. Liu, O. J. Kweon, Y. Xia, Z. Wu, and M. Chowdhury\n  The ML.ENERGY benchmark: toward automated inference energy measurement and optimization.\n  arXiv preprint arXiv:2505.06371.\n  Cited by: [Appendix A](#A1.p3.1 \"Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\"),\n  [§1](#S1.p1.1 \"1 Introduction ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\"),\n  [§3.2.1](#S3.SS2.SSS1.p2.1 \"3.2.1 Net value per dollar of AI cost and per AI hour ‣ 3.2 How do models trade off economic value and energy? (Q2) ‣ 3 Experiments and Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\"),\n  [§3.3](#S3.SS3.p4.1 \"3.3 How do accelerator type and count affect return per joule? (Q3) ‣ 3 Experiments and Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\").\n- CloudRift (2026)\n  CloudRift\n  GPU rental pricing and included resources.\n  Note: <https://www.cloudrift.ai/pricing>Accessed September 19, 2026\n  Cited by: [Appendix G](#A7.SS0.SSS0.Px1.p2.1 \"Rates and resource boundary. ‣ Appendix G Calculating AI-System Execution Cost ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\").\n- CoreWeave (2026)\n  CoreWeave\n  NVIDIA HGX H200 pricing.\n  Note: <https://www.coreweave.com/pricing>Accessed September 2026\n  Cited by: [§G.1](#A7.SS1.p1.2 \"G.1 Execution-cost sensitivity ‣ Appendix G Calculating AI-System Execution Cost ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\").\n- Deloitte (2026)\n  Deloitte\n  Deloitte’s enterprise AI infrastructure survey: a 2028 outlook.\n  Note: <https://www.deloitte.com/us/en/insights/topics/technology-management/ai-infrastructure-survey.html>Survey of 515 U.S. business and technology leaders, fielded November–December 2025\n  Cited by: [§1](#S1.p8.1 \"1 Introduction ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\").\n- Eloundou et al. (2023)\n  T. Eloundou, S. Manning, P. Mishkin, and D. Rock\n  GPTs are GPTs: an early look at the labor market impact potential of large language models.\n  arXiv preprint arXiv:2303.10130.\n  Cited by: [Appendix A](#A1.p2.1 \"Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\").\n- Erol et al. (2026)\n  M. H. Erol, B. El, M. Suzgun, M. Yuksekgonul, and J. Zou\n  Cost-of-pass: an economic framework for evaluating language models.\n  In International Conference on Learning Representations,\n  Cited by: [Appendix A](#A1.p3.1 \"Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\"),\n  [§1](#S1.p1.1 \"1 Introduction ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\").\n- Felten et al. (2021)\n  E. W. Felten, M. Raj, and R. Seamans\n  Occupational, industry, and geographic exposure to artificial intelligence: a novel dataset and its potential uses.\n  Strategic Management Journal 42 (12).\n  Cited by: [Appendix A](#A1.p2.1 \"Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\").\n- Google Cloud (2026)\n  Google Cloud\n  Accelerator-optimized machine family pricing.\n  Note: <https://cloud.google.com/products/compute/pricing/accelerator-optimized>Accessed September 20, 2026\n  Cited by: [§G.1](#A7.SS1.p1.2 \"G.1 Execution-cost sensitivity ‣ Appendix G Calculating AI-System Execution Cost ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\").\n- Google DeepMind (2026a)\n  Google DeepMind\n  Gemini 3.1 Flash-Lite model card.\n  Note: Official model documentation\n  Cited by: [§2.2](#S2.SS2.p1.1 \"2.2 Blinded evaluation ‣ 2 Methods ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\").\n- Google DeepMind (2026b)\n  Google DeepMind\n  Gemma 4 model card and Apache 2.0 license.\n  Note: <https://ai.google.dev/gemma/docs/core/model_card_4>Accessed September 2026\n  Cited by: [Appendix C](#A3.p2.1 \"Appendix C Benchmark Contamination ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\"),\n  [§3.2](#S3.SS2.p1.1 \"3.2 How do models trade off economic value and energy? (Q2) ‣ 3 Experiments and Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\").\n- Handa et al. (2025)\n  K. Handa A. Tamkin et al.\n  Which economic tasks are performed with AI? evidence from millions of Claude conversations.\n  arXiv preprint arXiv:2503.04761.\n  Cited by: [Appendix A](#A1.p2.1 \"Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\").\n- Hazy Research (2026a)\n  Hazy Research\n  Benchmarking intelligence efficiency of LM inference.\n  Note: <https://github.com/HazyResearch/intelligence-per-watt>Accessed September 2026\n  Cited by: [Appendix E](#A5.p2.1 \"Appendix E Execution and End-to-End Energy Accounting ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\").\n- Hazy Research (2026b)\n  Hazy Research\n  Telemetry overview.\n  Note: <https://hazyresearch.stanford.edu/intelligence-per-watt/benchmarking/overview/>Accessed September 22, 2026\n  Cited by: [Appendix E](#A5.p2.1 \"Appendix E Execution and End-to-End Energy Accounting ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\").\n- Jha et al. (2025)\n  S. Jha et al.\n  ITBench: evaluating AI agents across diverse real-world IT automation tasks.\n  In Proceedings of the 42nd International Conference on Machine Learning,\n  Proceedings of Machine Learning Research, Vol. 267.\n  Cited by: [Appendix A](#A1.p1.1 \"Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\").\n- Kapoor et al. (2024)\n  S. Kapoor et al.\n  AI agents that matter.\n  arXiv preprint arXiv:2407.01502.\n  Cited by: [Appendix A](#A1.p3.1 \"Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\").\n- Kapoor et al. (2025)\n  S. Kapoor, B. Stroebl, P. Kirgis, N. Nadgir, Z. S. Siegel, B. Wei, T. Xue, Z. Chen, et al.\n  Holistic agent leaderboard: the missing infrastructure for AI agent evaluation.\n  arXiv preprint arXiv:2510.11977.\n  Cited by: [Appendix A](#A1.p3.1 \"Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\").\n- Korinek and McKelvey (2026)\n  A. Korinek and P. McKelvey\n  Measuring the AI economy.\n  Technical report\n  Technical Report Staff Working Paper 2026-20, Bank of Canada.\n  Cited by: [Appendix A](#A1.p2.1 \"Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\").\n- Krupp et al. (2026)\n  L. Krupp, D. Geißler, V. Banwari, P. Lukowicz, and J. Karolus\n  Promoting sustainable web agents: benchmarking and estimating energy consumption through empirical and theoretical analysis.\n  Proceedings of the AAAI Conference on Artificial Intelligence 40 (45), pp. 38772–38780.\n  External Links: [Document](https://dx.doi.org/10.1609/aaai.v40i45.41221),\n  [Link](https://ojs.aaai.org/index.php/AAAI/article/view/41221)\n  Cited by: [Appendix A](#A1.p3.1 \"Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\").\n- Luccioni et al. (2024)\n  A. S. Luccioni, Y. Jernite, and E. Strubell\n  Power hungry processing: watts driving the cost of AI deployment?.\n  In Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency,\n  pp. 85–99.\n  Cited by: [Appendix A](#A1.p3.1 \"Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\").\n- Mazeika et al. (2025)\n  M. Mazeika, A. Gatti, C. Menghini, U. M. Sehwag, S. Singhal, Y. Orlovskiy, et al.\n  Remote labor index: measuring AI automation of remote work.\n  arXiv preprint arXiv:2510.26787.\n  Cited by: [Appendix A](#A1.p1.1 \"Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\").\n- Modal (2026)\n  Modal\n  Resource pricing: GPU, CPU and memory.\n  Note: <https://modal.com/pricing>Accessed September 19, 2026\n  Cited by: [Appendix G](#A7.SS0.SSS0.Px1.p2.1 \"Rates and resource boundary. ‣ Appendix G Calculating AI-System Execution Cost ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\").\n- National Renewable Energy Laboratory (2011)\n  National Renewable Energy Laboratory\n  Plug load assessment guidelines.\n  Technical report\n  Technical Report NREL/FS-7A20-50130, National Renewable Energy Laboratory.\n  External Links: [Link](https://docs.nlr.gov/docs/fy11osti/50130.pdf)\n  Cited by: [footnote 2](#footnote2 \"In 3.2 How do models trade off economic value and energy? (Q2) ‣ 3 Experiments and Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\").\n- Niu et al. (2026)\n  C. Niu, W. Zhang, J. Li, Y. Zhao, T. Wang, X. Wang, and Y. Chen\n  TokenPowerBench: benchmarking the power consumption of LLM inference.\n  Proceedings of the AAAI Conference on Artificial Intelligence 40 (38), pp. 32582–32590.\n  External Links: [Document](https://dx.doi.org/10.1609/aaai.v40i38.40535),\n  [Link](https://ojs.aaai.org/index.php/AAAI/article/view/40535)\n  Cited by: [Appendix A](#A1.SS0.SSS0.Px1.p1.1 \"From token efficiency to economic return. ‣ Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\").\n- Noy and Zhang (2023)\n  S. Noy and W. Zhang\n  Experimental evidence on the productivity effects of generative artificial intelligence.\n  Science 381 (6654), pp. 187–192.\n  Cited by: [Appendix A](#A1.p1.1 \"Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\").\n- NVIDIA (2026)\n  NVIDIA\n  NVIDIA H200 Tensor Core GPU: specifications.\n  Note: Official product specificationsH200 SXM maximum configurable TDP: 700 W. Accessed 2026-09-22\n  External Links: [Link](https://www.nvidia.com/en-us/data-center/h200/)\n  Cited by: [§3.2](#S3.SS2.p2.1 \"3.2 How do models trade off economic value and energy? (Q2) ‣ 3 Experiments and Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\").\n- O\\*NET Resource Center (2019)\n  O\\*NET Resource Center\n  O\\*NET-SOC 2019 taxonomy.\n  Note: National Center for O\\*NET Development\n  External Links: [Link](https://www.onetcenter.org/taxonomy/2019/)\n  Cited by: [Appendix H](#A8.p1.1 \"Appendix H Economic Weights and Wage Robustness ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\").\n- OpenAI (2025a)\n  OpenAI\n  gpt-oss model card and Apache 2.0 license.\n  Note: <https://huggingface.co/openai/gpt-oss-120b>Accessed September 2026\n  Cited by: [§3.2](#S3.SS2.p1.1 \"3.2 How do models trade off economic value and energy? (Q2) ‣ 3 Experiments and Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\").\n- OpenAI (2025b)\n  OpenAI\n  gpt-oss-120b & gpt-oss-20b model card.\n  Note: Official model documentation\n  Cited by: [Appendix C](#A3.p2.1 \"Appendix C Benchmark Contamination ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\").\n- OpenAI (2025c)\n  OpenAI\n  Measuring the performance of our models on real-world tasks.\n  Note: <https://openai.com/index/gdpval/>GDPval public announcement\n  Cited by: [Appendix C](#A3.p2.1 \"Appendix C Benchmark Contamination ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\").\n- Patwardhan et al. (2025)\n  T. Patwardhan, R. Dias, E. Proehl, G. Kim, M. Wang, O. Watkins, S. Posada Fishman, M. Aljubeh, P. Thacker, L. Fauconnet, N. S. Kim, P. Chao, S. Miserendino, G. Chabot, D. Li, M. Sharman, A. Barr, A. Glaese, and J. Tworek\n  GDPval: evaluating AI model performance on real-world economically valuable tasks.\n  arXiv preprint arXiv:2510.04374.\n  Cited by: [Appendix A](#A1.p1.1 \"Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\"),\n  [Appendix H](#A8.p1.1 \"Appendix H Economic Weights and Wage Robustness ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\"),\n  [§1](#S1.p1.1 \"1 Introduction ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\"),\n  [§2.1](#S2.SS1.p1.1 \"2.1 Dataset and expert references ‣ 2 Methods ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\"),\n  [§2.1](#S2.SS1.p4.1 \"2.1 Dataset and expert references ‣ 2 Methods ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\"),\n  [§2.2](#S2.SS2.p1.1 \"2.2 Blinded evaluation ‣ 2 Methods ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\").\n- Qwen Team (2026)\n  Qwen Team\n  Qwen3.6-27B model card and Apache 2.0 license.\n  Note: <https://huggingface.co/Qwen/Qwen3.6-27B>Accessed September 2026\n  Cited by: [Appendix C](#A3.p2.1 \"Appendix C Benchmark Contamination ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\"),\n  [§3.2](#S3.SS2.p1.1 \"3.2 How do models trade off economic value and energy? (Q2) ‣ 3 Experiments and Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\").\n- Saad-Falcon et al. (2025)\n  J. Saad-Falcon, A. Narayan, H. O. Akengin, J. W. Griffin, H. Shandilya, A. Gamarra Lafuente, M. Goel, R. Joseph, S. Natarajan, E. K. Guha, S. Zhu, B. Athiwaratkun, J. Hennessy, A. Mirhoseini, and C. Ré\n  Intelligence per watt: measuring intelligence efficiency of local AI.\n  arXiv preprint arXiv:2511.07885.\n  Cited by: [Appendix A](#A1.p3.1 \"Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\"),\n  [§2.3](#S2.SS3.p1.1 \"2.3 End-to-end energy accounting ‣ 2 Methods ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\").\n- Samsi et al. (2023)\n  S. Samsi, D. Zhao, J. McDonald, B. Li, A. Michaleas, M. Jones, W. Bergeron, J. Kepner, D. Tiwari, and V. Gadepally\n  From words to watts: benchmarking the energy costs of large language model inference.\n  arXiv preprint arXiv:2310.03003.\n  External Links: [Link](https://arxiv.org/abs/2310.03003)\n  Cited by: [Appendix A](#A1.SS0.SSS0.Px1.p1.1 \"From token efficiency to economic return. ‣ Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\").\n- Shao et al. (2025)\n  Y. Shao, H. Zope, Y. Jiang, J. Pei, D. Nguyen, E. Brynjolfsson, and D. Yang\n  Future of work with AI agents: auditing automation and augmentation potential across the U.S. workforce.\n  arXiv preprint arXiv:2506.06576.\n  Cited by: [Appendix A](#A1.p2.1 \"Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\").\n- Starace et al. (2025)\n  G. Starace et al.\n  PaperBench: evaluating AI’s ability to replicate AI research.\n  In Proceedings of the 42nd International Conference on Machine Learning,\n  Proceedings of Machine Learning Research, Vol. 267.\n  Cited by: [Appendix A](#A1.p1.1 \"Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\").\n- Tamkin and McCrory (2025)\n  A. Tamkin and P. McCrory\n  Estimating AI productivity gains from Claude conversations.\n  Note: Anthropic Economic Index\n  Cited by: [Appendix A](#A1.p2.1 \"Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\").\n- Tripathy et al. (2026)\n  A. Tripathy, C. P. Harshit, and K. Vaidhyanathan\n  SWEnergy: an empirical study on energy efficiency in agentic issue resolution frameworks with SLMs.\n  In Proceedings of the 2026 International Workshop on Agentic Engineering,\n  pp. 104–111.\n  External Links: [Document](https://dx.doi.org/10.1145/3786167.3788406),\n  [Link](https://doi.org/10.1145/3786167.3788406)\n  Cited by: [Appendix A](#A1.p3.1 \"Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\"),\n  [§1](#S1.p1.1 \"1 Introduction ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\").\n- Tschand et al. (2024)\n  A. Tschand, A. T. R. Rajan, S. Idgunji, A. Ghosh, J. Holleman, C. Kiraly, et al.\n  MLPerf Power: benchmarking the energy efficiency of machine learning systems from microwatts to megawatts for sustainable AI.\n  arXiv preprint arXiv:2410.12032.\n  Cited by: [Appendix A](#A1.SS0.SSS0.Px1.p1.1 \"From token efficiency to economic return. ‣ Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\"),\n  [Appendix A](#A1.p3.1 \"Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\").\n- U.S. Bureau of Labor Statistics (2025)\n  U.S. Bureau of Labor Statistics\n  May 2024 national occupational employment and wage estimates.\n  Note: Occupational Employment and Wage StatisticsMay 2024 estimates, released April 2025\n  External Links: [Link](https://www.bls.gov/oes/2024/may/oes_nat.htm)\n  Cited by: [Appendix H](#A8.p1.1 \"Appendix H Economic Weights and Wage Robustness ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\").\n- Vellaisamy et al. (2026)\n  P. Vellaisamy, V. Lam, S. Blanton, and J. P. Shen\n  Characterization of request and token energy costs for LLM inference workloads on GPU platforms.\n  arXiv preprint arXiv:2608.28044.\n  External Links: [Link](https://arxiv.org/abs/2608.28044)\n  Cited by: [Appendix A](#A1.SS0.SSS0.Px1.p1.1 \"From token efficiency to economic return. ‣ Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\").\n- Vidgen et al. (2026)\n  B. Vidgen, A. Mann, A. Fennelly, J. Wright Stanly, L. Rothman, M. Burstein, J. Benchek, D. Ostrofsky, A. Ravichandran, D. Sur, N. Venugopal, A. Hsia, I. Robinson, C. Huang, O. Varones, D. Khan, M. Haines, A. Bridges, J. Boyle, K. Twist, Z. Richards, C. Mahapatra, B. Foody, and O. Nitski\n  APEX–Agents.\n  Note: arXiv preprint arXiv:2601.14242\n  Cited by: [Appendix A](#A1.p1.1 \"Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\"),\n  [§K.2](#A11.SS2.p5.3 \"K.2 Repeated runs: Pass@𝑘 and four-run win–tie thresholds ‣ Appendix K Expert-Reference and Repeated-Run Sensitivity ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\").\n- Wan et al. (2026)\n  A. Wan, S. Hatgis-Kessell, T. Aguirre, P. Liang, and R. Bommasani\n  Economic evaluations of language models.\n  arXiv preprint arXiv:2607.19375.\n  Cited by: [Appendix A](#A1.p2.1 \"Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\"),\n  [§1](#S1.p1.1 \"1 Introduction ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\").\n- Wei et al. (2025)\n  T. Wei, W. Wen, R. Qiao, X. Sun, and J. Ma\n  RocketEval: efficient automated LLM evaluation via grading checklist.\n  In The Thirteenth International Conference on Learning Representations,\n  External Links: [Link](https://openreview.net/forum?id=zJjzNj6QUe)\n  Cited by: [§D.1](#A4.SS1.SSS0.Px2.p1.1 \"Expert review. ‣ D.1 Judge consistency ‣ Appendix D Judge and Evidence Pipeline ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\").\n- Xie et al. (2024)\n  T. Xie et al.\n  OSWorld: benchmarking multimodal agents for open-ended tasks in real computer environments.\n  arXiv preprint arXiv:2404.07972.\n  Cited by: [Appendix A](#A1.p1.1 \"Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\").\n- Z.ai (2026)\n  Z.ai\n  GLM-5.2 model card and MIT license.\n  Note: <https://huggingface.co/zai-org/GLM-5.2>Accessed September 2026\n  Cited by: [Appendix C](#A3.p2.1 \"Appendix C Benchmark Contamination ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\"),\n  [§3.2](#S3.SS2.p1.1 \"3.2 How do models trade off economic value and energy? (Q2) ‣ 3 Experiments and Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\").\n\n## Appendix A Related Work\n\nProfessional work and economic value. *GDPval*\n([Patwardhan et al., 2025](#bib.bib2)) evaluates deliverables;\n*APEX-Agents* and the *Remote Labor Index*\n([Vidgen et al., 2026](#bib.bib4); [Mazeika et al., 2025](#bib.bib8)) compare AI systems with experts\non longer tasks. Related benchmarks extend this approach to research, IT, web,\ndesktop, organizational settings, and economic simulations\n([Starace and others, 2025](#bib.bib25); [Jha and others, 2025](#bib.bib26); [Xie and others, 2024](#bib.bib27); [Backlund and Petersson, 2025](#bib.bib5)). Controlled and field studies complement these benchmarks\nby measuring changes in workers’ completion time, output quality, and productivity\n([Noy and Zhang, 2023](#bib.bib6); [Brynjolfsson et al., 2025](#bib.bib7)).\n\nOccupations and the wider economy. *GPTs Are GPTs*\n([Eloundou et al., 2023](#bib.bib28)) estimates which occupational tasks language models\ncould affect. The *Anthropic Economic Index*\n([Handa et al., 2025](#bib.bib29); [Tamkin and McCrory, 2025](#bib.bib9)) uses real Claude conversations\nto study where AI is being used at work and estimate possible time saving,\nwhile *Economic Evaluations of Language Models* ([Wan et al., 2026](#bib.bib30)) tests\ntasks across the U.S. occupational economy and estimates how much work time current\nmodels could save. Related studies connect AI capabilities, use, and adoption\nto workers, occupations, firms, industries, and the broader economy\n([Felten et al., 2021](#bib.bib31); [Shao et al., 2025](#bib.bib10); [Bonney et al., 2026](#bib.bib11); [Acemoglu, 2024](#bib.bib13); [Korinek and McKelvey, 2026](#bib.bib12)).\n\nCost and energy efficiency. *AI Agents\nThat Matter*, *Cost-of-Pass*, and the *Holistic Agent Leaderboard* compare\nperformance with financial or usage-based costs\n([Kapoor and others, 2024](#bib.bib32); [Erol et al., 2026](#bib.bib14); [Kapoor et al., 2025](#bib.bib15)). Energy studies such as\n*Power Hungry Processing*, *ML.ENERGY*, and *MLPerf Power* directly measure energy\nuse across models, workloads, and accelerators\n([Luccioni et al., 2024](#bib.bib17); [Chung et al., 2025](#bib.bib16); [Tschand et al., 2024](#bib.bib18); [Saad-Falcon et al., 2025](#bib.bib19)). More recent\nwork, including *SWEnergy* and *Sustainable Web Agents*, extends energy measurement\nto AI-system workflows ([Tripathy et al., 2026](#bib.bib33); [Krupp et al., 2026](#bib.bib34)).\nNVJ and TSJ extend this resource accounting to net dollar value and time saving\nrather than benchmark scores or win–tie rates alone.\n\n##### From token efficiency to economic return.\n\nToken-level studies vary batch size, context length, and GPU count, including\nphase-specific prefill and decoding measurements\n([Samsi et al., 2023](#bib.bib49); [Niu et al., 2026](#bib.bib50)). Longer generations can lower\njoules per token by amortizing prefill while increasing total request energy\n([Vellaisamy et al., 2026](#bib.bib51)); long contexts enlarge attention and\nkey–value-cache traffic. For a common workload, window, and energy boundary,\ntokens per joule multiplied by net value or time saving per output token\ngives NVJ or TSJ, after unit conversion. Token efficiency therefore complements\noutput quality, task value, and time saving; tool execution and waiting also\nconsume energy without producing model tokens. Comparing these measures requires\nmatched boundaries because GPU counters and MLPerf Power’s external whole-system\nmetering cover different components ([Tschand et al., 2024](#bib.bib18)).\n\n## Appendix B Expert References and Annotations\n\nThe tasks span heterogeneous formats, including documents,\nspreadsheets, presentations, code, images, audio, video, and archive-based\ndeliverables. The frozen task records also preserve occupation, sector, and\nO\\*NET-SOC identifiers; source-file and deliverable inventories; and the\ngrading rubric. The expert-only reference was produced by an expert without\nAI; its recorded time is decomposed into prompt review,\nsource-file use, task execution, and verification or quality review.\nThe expert-with-AI reference was produced by an expert using AI. Its records\nadd the AI provider, model version, interface, source files\nused, external tools, full transcripts, and turn-level\nexpert-active, AI-generation-wait, and total times.\nThe supplied assisted-reference records identify Claude Sonnet 4.6 and\nClaude Opus 4.7 used through Claude Web Chat.\nThese are recorded expert completion times. The assisted condition includes\nthe expert’s setup, monitoring, review, and correction, as well as AI waits.\nThe expert-only and expert-with-AI references were annotated by different people.\n\nThe pairwise comparison with an expert deliverable assesses relative judged\nquality; it does not establish absolute task quality or realized economic\nsavings. We use the GDPval expert deliverables as meaningful reference points\nwithout assuming they are error-free.\n\nThe rubric helps ground the\ncomparison in the task requirements even when an expert deliverable has flaws.\nExpert-only reference, expert-with-AI reference, and\nAI-system deliverables are normalized to the same evidence representation.\nCandidate position is balanced exactly\nacross the archived task-condition comparisons.\n\nThe expert-with-AI reference is a supplied assisted completion condition;\nit is not a claim about every possible form of assisted professional work.\nIts completion time can exceed expert-only time on individual tasks.\nThe expert-only reference is the cleaner\ncommon reference for autonomous systems because neither its\ndeliverables nor its output dollar values depend on a particular assisting model,\ninterface, or expert–AI workflow. We therefore retain it as primary; the\nexpert-with-AI reference is a secondary, workflow-specific\noperational comparison. Appendix [K.1](#A11.SS1 \"K.1 Separating reference judgments from time valuation ‣ Appendix K Expert-Reference and Repeated-Run Sensitivity ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\") separates\nchanges in judged win–tie outcomes from changes in time valuation.\n\n## Appendix C Benchmark Contamination\n\nDuring evaluation, the models received each task and its required source files,\nbut not the grading rubric or expert deliverable. We checked AI deliverables\nfor exact matches to expert files and unexplained identical document or\nworkbook components and found none.\n\nGPT-OSS 120B and GPT-OSS 20B were released\nbefore GDPval became public ([OpenAI, 2025b](#bib.bib23); [OpenAI, 2025c](#bib.bib3)).\nThe public model cards for GLM 5.2, Qwen3.6-27B, and Gemma 4 31B do not\nmention GDPval\n([Z.ai, 2026](#bib.bib44); [Qwen Team, 2026](#bib.bib45); [Google DeepMind, 2026b](#bib.bib46)).\nThese checks show no direct copying in the evaluated outputs; they do not\nestablish every model’s training-data contents.\n\n## Appendix D Judge and Evidence Pipeline\n\nThe judging pipeline is model-agnostic and can be configured with\nalternative language-model judges. Main economic results use the\nfrozen judge models/gemini-3.1-flash-lite at temperature 0. Its input\ncontains the task specification, grading rubric, source evidence, and both\nidentity-masked candidate deliverables. The evidence pipeline uses full sanitized\nsemantic text and rendered visual PDFs with a 48-page visual ceiling. It\nperforms safe archive extraction, static code inspection, and media parsing.\nAll verdict records used in the analysis parsed without error.\n\n### D.1 Judge consistency\n\nTable [2](#A4.T2 \"Table 2 ‣ D.1 Judge consistency ‣ Appendix D Judge and Evidence Pipeline ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\") compares Gemini 3.1 Flash-Lite (the main\njudge), Claude Haiku 4.5, and GPT-5 mini on the same GLM 5.2 comparisons,\nincluding all four attempts per task.\n\nTable 2: The three judges agree on 72–77% of exact outcomes. An AI win or tie receives task credit. Exact agreement\nretains win/tie/loss; binary win–tie agreement collapses wins and ties into one category. All comparisons use GLM 5.2 outputs from four attempts per task.\n\n| Pair | Exact agreement | Binary win–tie agreement |\n| --- | --- | --- |\n| Gemini–Haiku | 77.2% | 79.2% |\n| Gemini–GPT-5 mini | 74.0% | 76.4% |\n| Haiku–GPT-5 mini | 72.3% | 72.8% |\n\n##### Repeated-output agreement.\n\nFor GLM 5.2, the first two independently generated outputs for each of 220\ntasks produced 440 Gemini judgments. The win–tie versus loss decision matched\nacross the two outputs for 171 tasks (77.7%).\n\n##### Expert review.\n\nIn a blinded expert review of 32 randomly selected GLM 5.2 task comparisons,\nexperts agreed with Gemini 3.1 Flash-Lite on 23 of 32 win–tie versus loss\ndecisions (71.9%). A separate automated review yielded the same agreement\nrate. Human-to-human agreement on related pairwise response judgments has\nbeen reported at 64.7% ([Wei et al., 2025](#bib.bib1)).\n\n## Appendix E Execution and End-to-End Energy Accounting\n\nThe AI-system runs were executed in concurrent batches. IPW telemetry\nsampled the metered components at a configured 0.2-second\ninterval. The end-to-end measurement window began immediately before the\nbatch’s runs launched and ended when the final run completed. For\ncomponents exposing cumulative energy counters, end-to-end energy used the\nboundary-aligned difference between the end and start readings. Otherwise,\nsampled power was integrated over the same wall-clock interval. Measurements\nfrom all accelerators assigned to a run were summed over the execution.\n\nFor every evaluated deployment, IPW measured GPU, CPU, and DRAM energy.\nReported joules sum these component-energy readings over the execution\nwindow ([Hazy Research, 2026a](#bib.bib48); [Hazy Research, 2026b](#bib.bib52)).\nTable [3](#A5.T3 \"Table 3 ‣ Appendix E Execution and End-to-End Energy Accounting ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\") gives the serving settings.\n\nTable 3: All evaluated deployments use vLLM, 20 concurrent requests, a 131,072-token context limit, and a 32,768-token output limit. Tensor parallelism is eight for eight-accelerator deployments and one for single-accelerator deployments.\n\n| Model and accelerator | Inference | Tensor parallelism | Concurrent requests | Context limit (tokens) | Output limit (tokens) |\n| --- | --- | --- | --- | --- | --- |\n| GLM 5.2, 8×\\timesH200 | vLLM | TP8 | 20 | 131,072 | 32,768 |\n| Qwen3.6-27B, 8×\\timesH100 | vLLM | TP8 | 20 | 131,072 | 32,768 |\n| Gemma 4 31B, 8×\\timesH100 | vLLM | TP8 | 20 | 131,072 | 32,768 |\n| GPT-OSS 120B, 8×\\timesH100 | vLLM | TP8 | 20 | 131,072 | 32,768 |\n| GPT-OSS 20B, 8×\\timesH100 | vLLM | TP8 | 20 | 131,072 | 32,768 |\n| Qwen3.6-27B, 1×\\timesB200 | vLLM | TP1 | 20 | 131,072 | 32,768 |\n| Qwen3.6-27B, 1×\\timesH200 | vLLM | TP1 | 20 | 131,072 | 32,768 |\n| Qwen3.6-27B, 1×\\timesH100 | vLLM | TP1 | 20 | 131,072 | 32,768 |\n| Qwen3.6-27B, 1×\\timesMI350X | vLLM | TP1 | 20 | 131,072 | 32,768 |\n\nWe chose OpenHands as the primary harness because its compact tool surface, a\nterminal and file editor, was sufficient for producing and validating the\nheterogeneous deliverables in GDPval. The run ledger records turns, tool\ncalls, AI-system completion time, outcome, and energy, allowing the same evaluation protocol to\nbe applied to other harnesses and inference backends. The codebase also includes\nportability paths for MiniMax and Apple M4; these\nconditions are not pooled with the main results unless the complete evaluation\nprotocol is matched.\n\nThe active-request allocation in Equation [2](#S2.E2 \"In 2.3 End-to-end energy accounting ‣ 2 Methods ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\") reconciles\nto the retained batch totals. This conservation identity applies to a full\nbatch; it does not imply that concurrent requests draw equal instantaneous\npower.\n\n## Appendix F Energy-Attribution Sensitivity\n\nA concurrent batch has one measured energy total, while individual requests\noverlap in time. Our primary active-request rule\n(Equation [2](#S2.E2 \"In 2.3 End-to-end energy accounting ‣ 2 Methods ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\")) assigns each moment’s energy across the\nrequests then running. An equal-batch-share alternative assigns the same\nfraction of the batch total to every request. Both rules conserve measured\nbatch energy, but they can assign different energy to individual tasks\nbecause requests finish at different times.\n\nWe apply both rules to the evaluated GLM 5.2 task pool underlying\nFigure [3](#S3.F3 \"Figure 3 ‣ 3 Experiments and Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\"), holding AI outputs, judge decisions,\nexpert times, wages, AI-system execution costs, and GDP weights fixed.\nTable [4](#A6.T4 \"Table 4 ‣ Appendix F Energy-Attribution Sensitivity ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\") gives the category comparisons.\nProtective Service remains first in win–tie rate per MJ, and Computer &\nMathematical remains first in NVJ. The TSJ leader changes. Architecture & Engineering leads\nunder the primary active-request rule, while Computer & Mathematical leads\nunder equal batch shares. Eight NVJ ranks and nine TSJ ranks change across\nthe 13 categories. Both rules conserve each measured batch total; neither establishes\nthe energy drawn by an individual concurrent request.\n\nFigure 8: Request-completion share does not track batch energy proportionally. P20, P50, P80, and P90 mark completion of 20%, 50%, 80%, and 90% of requests. Each point pools batch totals after estimating energy consumed at that milestone. Intermediate points use each batch’s average power and recorded completion times; the 100% endpoint is measured batch energy. The dotted line shows proportional completion for reference.\n\nTable 4: Equal batch shares change the TSJ leader but leave the NVJ leader unchanged. Each cell shows active-request allocation / equal batch share. Active-request allocation is the paper’s primary method. AI outputs, judgments, expert times, wages, costs, and GDP weights are fixed; only energy assignment changes. Values are point estimates.\n\n| Category | Energy (MJ) | NVJ (USD/MJ) | TSJ (hours/MJ) |\n| --- | --- | --- | --- |\n| Computer & Mathematical | 0.355 / 0.337 | 523.8 / 551.6 | 7.29 / 7.67 |\n| Architecture & Engineering | 0.495 / 0.597 | 413.9 / 343.2 | 7.89 / 6.54 |\n| Management | 0.360 / 0.349 | 374.6 / 387.0 | 6.75 / 6.97 |\n| Protective Service | 0.225 / 0.229 | 278.1 / 272.9 | 4.96 / 4.86 |\n| Production | 0.550 / 0.786 | 221.1 / 154.6 | 6.12 / 4.28 |\n| Community & Social Service | 0.282 / 0.363 | 188.5 / 146.3 | 6.16 / 4.78 |\n| Legal | 0.370 / 0.334 | 149.7 / 166.1 | 1.50 / 1.66 |\n| Business & Financial Operations | 0.505 / 0.481 | 145.6 / 152.9 | 3.32 / 3.49 |\n| Personal Care & Service | 0.401 / 0.503 | 134.3 / 107.2 | 7.15 / 5.70 |\n| Arts, Design, Entertainment & Media | 0.450 / 0.427 | 103.8 / 109.5 | 2.71 / 2.86 |\n| Healthcare Practitioners & Technical | 0.445 / 0.504 | 97.8 / 86.3 | 1.67 / 1.47 |\n| Sales & Related | 0.356 / 0.362 | 65.4 / 64.4 | 2.01 / 1.98 |\n| Office & Administrative Support | 0.604 / 0.567 | 40.1 / 42.6 | 1.84 / 1.96 |\n\nFigure [8](#A6.F8 \"Figure 8 ‣ Appendix F Energy-Attribution Sensitivity ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\") describes when energy is consumed during\nthe measured execution batches. For each batch, we find when a given fraction\nof its requests has finished and multiply elapsed batch time by its measured\naverage power. P20, P50, P80, and P90 correspond to an estimated 37.9%,\n52.4%, 71.1%, and 83.9% of pooled batch energy. These intermediate values\nuse a constant-power approximation; only the full-batch energy total is\ndirectly measured. The figure does not evaluate a stopping policy, since\nenergy already spent on unfinished requests is included at each milestone\nand the outcomes of stopping them are not observed.\n\n## Appendix G Calculating AI-System Execution Cost\n\nWe calculate AI-system execution cost for the evaluated runs from recorded\nGPU, CPU, and RAM use at published rates. Those rates include hardware and\nelectricity, so energy is not charged again in dollars.\n\n##### Rates and resource boundary.\n\nThe eight-H100 comparisons use the July 2026 AWS Linux On-Demand\np5.48xlarge rate of $55.04 per node-hour. GLM 5.2 uses the\np5en.48xlarge rate of $63.296 per node-hour for eight H200s.\nThese are whole-node rates including host CPU and RAM, not rates per GPU.\nThe calculation also includes one boot volume, one public IPv4 address, and\none final artifact export where applicable\n([Amazon Web Services, 2026d](#bib.bib35); [Amazon Web Services, 2026c](#bib.bib36); [Amazon Web Services, 2026a](#bib.bib37); [Amazon Web Services, 2026e](#bib.bib38); [Amazon Web Services, 2026b](#bib.bib39)).\n\nFor the single-NVIDIA-accelerator comparison, we use Modal’s published\nper-second H100, H200, and B200 prices together with a common 32-core,\n256-GiB host reservation ([Modal, 2026](#bib.bib40)). The MI350X comparison\nuses CloudRift’s published on-demand GPU-hour rate\n([CloudRift, 2026](#bib.bib43)). These provider prices do not guarantee\nidentical host specifications or performance across providers.\n\n##### Time allocation and setup.\n\nLet Ai​(t)A\\_{i}(t) indicate when request ii is active, and let N⁡(t)N(t)\ncount active requests. We allocate shared active node time using\nτi=∫Ai​(t)/N⁡(t)​𝑑t\\tau\\_{i}=\\int A\\_{i}(t)/N(t)\\,dt. Shared startup, gaps, and finalization\ntime T0T\\_{0} inside a task-execution window contributes\nT0​τi/∑jτjT\\_{0}\\tau\\_{i}/\\sum\\_{j}\\tau\\_{j}, so the recorded window is charged once.\nThus the total node time allocated to request ii is\nTialloc=τi+T0​τi/∑jτjT\\_{i}^{\\mathrm{alloc}}=\\tau\\_{i}+T\\_{0}\\tau\\_{i}/\\sum\\_{j}\\tau\\_{j}.\nThe comparison represents a prepared local service; initialization outside\nthe task-execution window is excluded from the reported execution cost.\n\n##### Scope.\n\nPublished prices can change. The calculation excludes unobserved\nexternal API charges, durable storage, expert review, development,\ncoordination, and adoption costs. These belong in a broader deployment\nassessment alongside the measured energy and the value of delivered work.\n\n### G.1 Execution-cost sensitivity\n\nFor each evaluated GLM 5.2 task, we calculate the AI-system execution cost by\nmultiplying its allocated node seconds by the AWS node-hour rate divided by\n3,600, then adding its assigned boot-volume and IPv4 charges. To examine a\nhypothetical alternative provider price, we replace only that hourly node\nrate. For task (i), the resulting cost at rate (r) is\n\n|  |  |  |  |\n| --- | --- | --- | --- |\n|  | Ci​(r)=Ci​(AWS)+Tialloc3,600​(r−63.296),C\\_{i}(r)=C\\_{i}(\\mathrm{AWS})+\\frac{T\\_{i}^{\\mathrm{alloc}}}{3{,}600}(r-63.296), |  | (6) |\n\nwhere TiallocT\\_{i}^{\\mathrm{alloc}} includes active and allocated shared node seconds. We use published eight-H200\nnode tariffs of $50.44/hour for CoreWeave and $84.8069/hour for Google Cloud\nalongside the $63.296/hour AWS rate\n([CoreWeave, 2026](#bib.bib42); [Google Cloud, 2026](#bib.bib41); [Amazon Web Services, 2026d](#bib.bib35)).\nThe same recorded outcomes, time, energy, and non-node charges are held fixed;\nthe alternative-provider columns are hypothetical price calculations, not\nruns on those providers. Table [5](#A7.T5 \"Table 5 ‣ G.1 Execution-cost sensitivity ‣ Appendix G Calculating AI-System Execution Cost ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\") shows\nthe resulting category costs and NVJ. The category NVJ order is unchanged\nat these three tariffs, while TSJ is unchanged because execution cost does\nnot enter its numerator or denominator.\n\nTable 5: Published node tariffs change calculated costs without changing category NVJ order. For the same recorded GLM 5.2 tasks, only the node-hour price is changed; task outcomes, time, energy, and non-node charges remain fixed. Values are hypothetical costs per category-normalized task and the resulting NVJ, not results of executions on other providers. The AWS column reproduces the evaluated cost boundary.\n\n|  | AI execution cost (USD) | | | NVJ (USD/MJ) | |\n| --- | --- | --- | --- | --- | --- |\n| Category | CoreWeave | AWS | GCP | CoreWeave | GCP |\n| Architecture & Eng. | 1.677 | 2.104 | 2.819 | 414.8 | 412.5 |\n| Arts & Media | 1.582 | 1.986 | 2.660 | 104.7 | 102.3 |\n| Business & Financial | 1.731 | 2.172 | 2.911 | 146.4 | 144.1 |\n| Community & Social | 0.983 | 1.233 | 1.652 | 189.4 | 187.0 |\n| Computer & Math | 1.245 | 1.562 | 2.092 | 524.6 | 522.3 |\n| Healthcare | 1.837 | 2.306 | 3.089 | 98.9 | 96.1 |\n| Legal | 1.257 | 1.577 | 2.113 | 150.6 | 148.3 |\n| Management | 1.256 | 1.576 | 2.112 | 375.5 | 373.1 |\n| Office & Admin. | 2.173 | 2.727 | 3.653 | 41.0 | 38.5 |\n| Personal Care | 1.506 | 1.890 | 2.532 | 135.3 | 132.7 |\n| Production | 1.859 | 2.333 | 3.125 | 222.0 | 219.7 |\n| Protective Service | 0.778 | 0.976 | 1.308 | 279.0 | 276.6 |\n| Sales | 1.265 | 1.587 | 2.127 | 66.3 | 63.9 |\n\n## Appendix H Economic Weights and Wage Robustness\n\nGDPval chose sectors for their GDP contribution and sampled predominantly\nknowledge-work occupations with large aggregate compensation ([Patwardhan et al., 2025](#bib.bib2)).\nIts public gold set has five tasks per occupation, so equal task weights would\nreflect the sampling design rather than sector economic scale. We give each of\nthe nine sectors its normalized GDP share, then divide that share among its\nevaluated occupations in proportion to national employment. Sector GDP thus\nsets the scale across sectors, while employment distinguishes occupations\nwithin each sector. We do not use wage-bill weights because wages already enter\noutput dollar value as recorded expert completion time multiplied by the\noccupation’s matched OEWS median hourly wage ([U.S. Bureau of Labor Statistics, 2025](#bib.bib21); [Patwardhan et al., 2025](#bib.bib2)).\nUsing wages again in the weights would give high-wage occupations additional\ninfluence; wage-bill weights would instead emphasize compensation paid to those\noccupations. The GDP-based weights preserve sector GDP shares without assuming\nequal GDP-to-wage-bill ratios across sectors. The 44 occupations are grouped into 13\noccupation categories using the O\\*NET-SOC taxonomy ([O\\*NET Resource Center, 2019](#bib.bib22));\nO\\*NET-SOC supplies the classification and does not construct the weights.\n\nLet γs\\gamma\\_{s} be sector ss’s share of U.S. GDP, normalized across the\nnine selected sectors, and let\nLo=EmploymentoL\\_{o}=\\mathrm{Employment}\\_{o}. The occupation\nweight used in both aggregation levels is\n\n|  |  |  |  |\n| --- | --- | --- | --- |\n|  | woGDP=γs⁡(o)Lo∑j:s⁡(j)=s⁡(o)Lj,∑o∈𝒪woGDP=1.w\\_{o}^{\\mathrm{GDP}}=\\gamma\\_{s(o)}\\frac{L\\_{o}}{\\sum\\_{j:s(j)=s(o)}L\\_{j}},\\qquad\\sum\\_{o\\in\\mathcal{O}}w\\_{o}^{\\mathrm{GDP}}=1. |  | (7) |\n\nFor completeness, define the per-occupation task average as\nX~o=Xo/no\\widetilde{X}\\_{o}=X\\_{o}/n\\_{o}, where non\\_{o} is its evaluated task count and\nX∈{G,C,S,E}X\\in\\{G,C,S,E\\}, with SS denoting time saving as in\nSection [2.5](#S2.SS5 \"2.5 Two levels of GDP-based aggregation ‣ 2 Methods ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\"). For category cc, its category weight and within-category\noccupation weights are\n\n|  |  |  |  |\n| --- | --- | --- | --- |\n|  | Wc=∑o∈cwoGDP,wo|c=woGDPWc,∑o∈cwo|c=1.W\\_{c}=\\sum\\_{o\\in c}w\\_{o}^{\\mathrm{GDP}},\\qquad w\\_{o\\mid c}=\\frac{w\\_{o}^{\\mathrm{GDP}}}{W\\_{c}},\\qquad\\sum\\_{o\\in c}w\\_{o\\mid c}=1. |  | (8) |\n\nThe category outcome is\nX~c=∑o∈cwo|c​X~o\\widetilde{X}\\_{c}=\\sum\\_{o\\in c}w\\_{o\\mid c}\\widetilde{X}\\_{o}, while the\nGDP-weighted benchmark outcome is\nX~GDP=∑cWc​X~c\\widetilde{X}^{\\mathrm{GDP}}=\\sum\\_{c}W\\_{c}\\widetilde{X}\\_{c}, with\n∑cWc=1\\sum\\_{c}W\\_{c}=1. The same within-category and benchmark weights are applied to\noutput dollar value, AI-system execution cost, time saving and energy before\nforming NVJ, TSJ or net value per dollar of AI cost; the method never\naverages already-formed occupation or category ratios.\n\nTable [6](#A8.T6 \"Table 6 ‣ Appendix H Economic Weights and Wage Robustness ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\") lists the resulting category weights, and\nFigure [9](#A8.F9 \"Figure 9 ‣ Appendix H Economic Weights and Wage Robustness ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\") visualizes their construction and use.\n\nFigure 9: Wages determine output dollar value without also increasing benchmark weights. Sector GDP shares and within-sector employment shares determine occupation weights. Category comparisons renormalize those weights to one within each category, while the GDP-weighted benchmark metric retains category weights. Expert completion time and occupation-specific median wages determine output dollar value.\n\nTable 6: Sales and Management receive the largest GDP-based category weights. Employment shares allocate each sector’s GDP share among its evaluated occupations; the category weights sum to one and are retained in the GDP-weighted benchmark aggregate.\n\nCategory\nOccupations\nWeight\nEmployment (M)\n\nManagement\n7\n19.68%\n6.77\n\nBusiness & Financial Ops.\n6\n12.44%\n3.95\n\nComputer & Mathematical\n1\n3.22%\n1.65\n\nArchitecture & Engineering\n2\n3.16%\n0.64\n\nCommunity & Social Service\n1\n3.82%\n0.38\n\nLegal\n1\n1.45%\n0.75\n\nArts, Design, Entertainment & Media\n5\n7.13%\n0.38\n\nHealthcare Practitioners & Technical\n3\n6.10%\n3.92\n\nProtective Service\n2\n1.59%\n0.19\n\nPersonal Care & Service\n2\n3.91%\n0.35\n\nSales & Related\n8\n20.24%\n4.00\n\nOffice & Administrative Support\n5\n13.87%\n5.99\n\nProduction\n1\n3.39%\n0.69\n\nTable [7](#A8.T7 \"Table 7 ‣ Appendix H Economic Weights and Wage Robustness ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\") compares mean and median hourly wages in task\nvaluation while holding GDP-based occupation weights, energy, AI-system execution cost, and\ntime saving fixed. The headline analysis uses median hourly wages.\nU.S. GDP shares, employment, wages, work practices, and estimates of time saving\nchange over time; results should therefore be recomputed as work changes.\n\nTable 7: Mean rather than median wages raise NVJ while leaving TSJ unchanged. AI-system execution cost is held fixed for GDP-weighted GLM 5.2 with one attempt per task. NVJ is USD/MJ; TSJ is hours/MJ.\n\nWage basis\n\nOutput dollar\nvalue (USD)\n\nAI-system\nexecution cost (USD)\nEnergy (MJ)\nNVJ\nTSJ\n\nOEWS mean\n85.57\n1.92\n0.430\n194.58\n3.823\n\nOEWS median\n74.19\n1.92\n0.430\n168.11\n3.823\n\n## Appendix I Occupation-Category Results\n\nTable [8](#A9.T8 \"Table 8 ‣ Appendix I Occupation-Category Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\") gives win–tie rates,\nTable [9](#A9.T9 \"Table 9 ‣ Appendix I Occupation-Category Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\") gives net dollar value and time saving\nper task, and Table [10](#A9.T10 \"Table 10 ‣ Appendix I Occupation-Category Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\") gives returns per MJ.\nThese tables underlie Figure [3](#S3.F3 \"Figure 3 ‣ 3 Experiments and Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\"). Brackets give 95% task-composition\nintervals. They describe variation across the evaluated tasks and do not\nestablish a population ordering of categories.\n\nTable 8: Protective Service leads win–tie rate per MJ, while Personal Care leads win–tie rate. GLM 5.2, expert-only reference, one attempt per task; occupation outcomes are category-normalized with GDP-based weights. Brackets show 95% task-composition intervals.\n\n| Category | Win–tie (%) | Win–tie pp/MJ |\n| --- | --- | --- |\n| Architecture & Eng. | 69.00 [38.01, 91.00] | 139.43 [75.84, 229.50] |\n| Arts & Media | 46.09 [28.27, 66.09] | 102.44 [74.15, 136.56] |\n| Business & Financial | 56.81 [37.93, 74.97] | 112.42 [72.81, 168.62] |\n| Community & Social | 60.00 [20.00, 100.00] | 212.94 [60.19, 390.26] |\n| Computer & Math | 80.00 [40.00, 100.00] | 225.08 [122.22, 312.77] |\n| Healthcare | 28.23 [8.02, 63.12] | 63.48 [21.77, 107.89] |\n| Legal | 60.00 [20.00, 100.00] | 162.14 [49.98, 321.76] |\n| Management | 78.67 [62.10, 92.17] | 218.37 [162.96, 289.27] |\n| Office & Admin. | 38.14 [16.10, 62.36] | 63.18 [23.59, 135.67] |\n| Personal Care | 95.79 [87.37, 100.00] | 238.63 [150.53, 515.22] |\n| Production | 40.00 [0.00, 80.00] | 72.78 [0.00, 150.11] |\n| Protective Service | 60.00 [20.80, 99.20] | 267.01 [84.52, 479.38] |\n| Sales | 55.31 [34.84, 74.31] | 155.28 [96.17, 222.56] |\n\nTable 9: Architecture & Engineering leads net dollar value and time saving per task. GLM 5.2, expert-only reference, one attempt per task; occupation outcomes are category-normalized with GDP-based weights. Brackets show 95% task-composition intervals.\n\n| Category | Net dollar value (USD/task) | Time saving (hours/task) |\n| --- | --- | --- |\n| Architecture & Eng. | 204.84 [108.93, 288.94] | 3.905 [2.055, 5.569] |\n| Arts & Media | 46.71 [20.18, 79.15] | 1.219 [0.554, 2.028] |\n| Business & Financial | 73.57 [44.46, 106.85] | 1.677 [0.975, 2.508] |\n| Community & Social | 53.12 [12.78, 93.45] | 1.735 [0.423, 3.047] |\n| Computer & Math | 186.16 [68.11, 337.13] | 2.589 [0.924, 4.844] |\n| Healthcare | 43.50 [14.21, 86.93] | 0.742 [0.230, 1.541] |\n| Legal | 55.40 [7.06, 117.50] | 0.555 [0.046, 1.338] |\n| Management | 134.95 [73.68, 229.83] | 2.431 [1.239, 4.337] |\n| Office & Admin. | 24.18 [5.74, 48.62] | 1.111 [0.315, 2.200] |\n| Personal Care | 53.92 [34.42, 73.33] | 2.870 [1.753, 3.975] |\n| Production | 121.51 [-2.22, 250.73] | 3.361 [0.000, 6.915] |\n| Protective Service | 62.49 [7.21, 150.74] | 1.114 [0.128, 2.792] |\n| Sales | 23.29 [15.22, 31.58] | 0.717 [0.455, 0.987] |\n\nTable 10: Computer & Mathematical leads NVJ; Architecture & Engineering leads TSJ. GLM 5.2, expert-only reference, one attempt per task; occupation outcomes are category-normalized with GDP-based weights. Brackets show 95% task-composition intervals.\n\n| Category | NVJ (USD/MJ) | TSJ (hours/MJ) |\n| --- | --- | --- |\n| Architecture & Eng. | 413.90 [196.34, 744.61] | 7.890 [3.660, 14.397] |\n| Arts & Media | 103.82 [55.09, 167.63] | 2.710 [1.459, 4.508] |\n| Business & Financial | 145.57 [83.22, 239.58] | 3.318 [1.820, 5.627] |\n| Community & Social | 188.50 [43.10, 382.42] | 6.158 [1.426, 12.522] |\n| Computer & Math | 523.76 [235.93, 770.14] | 7.285 [3.182, 11.080] |\n| Healthcare | 97.84 [38.77, 196.72] | 1.669 [0.643, 3.049] |\n| Legal | 149.70 [21.88, 370.36] | 1.500 [0.144, 4.257] |\n| Management | 374.57 [193.82, 672.56] | 6.748 [3.253, 12.692] |\n| Office & Admin. | 40.05 [8.22, 105.52] | 1.840 [0.470, 4.745] |\n| Personal Care | 134.32 [87.59, 261.20] | 7.151 [4.532, 14.139] |\n| Production | 221.10 [-4.24, 479.79] | 6.116 [0.000, 13.337] |\n| Protective Service | 278.09 [30.55, 669.23] | 4.956 [0.533, 12.268] |\n| Sales | 65.40 [42.16, 93.84] | 2.014 [1.283, 2.898] |\n\nTable [11](#A9.T11 \"Table 11 ‣ Appendix I Occupation-Category Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\") reports output dollar value and time saving\nper win or tie. These components explain why a category can yield fewer wins\nand ties per joule while delivering greater dollar value or time saving per\njoule.\n\nTable 11: Win–tie values and expert times explain why the category rankings differ. Values condition on outputs judged a win or tie for GLM 5.2 against the expert-only reference. Occupation outcomes are category-normalized with GDP-based weights.\n\n| Category | Output dollar value (USD) | Expert time (h) | AI time (h) | Time saving (h) |\n| --- | --- | --- | --- | --- |\n| Architecture & Eng. | 299.90 | 6.11 | 0.45 | 5.66 |\n| Arts & Media | 105.66 | 3.02 | 0.37 | 2.65 |\n| Business & Financial | 133.31 | 3.34 | 0.39 | 2.95 |\n| Community & Social | 90.58 | 3.22 | 0.32 | 2.89 |\n| Computer & Math | 234.65 | 3.67 | 0.43 | 3.24 |\n| Healthcare | 162.29 | 3.04 | 0.41 | 2.63 |\n| Legal | 94.96 | 1.31 | 0.38 | 0.93 |\n| Management | 173.53 | 3.39 | 0.30 | 3.09 |\n| Office & Admin. | 70.55 | 3.30 | 0.38 | 2.91 |\n| Personal Care | 58.26 | 3.40 | 0.40 | 3.00 |\n| Production | 309.61 | 9.04 | 0.64 | 8.40 |\n| Protective Service | 105.78 | 2.13 | 0.27 | 1.86 |\n| Sales | 44.98 | 1.63 | 0.34 | 1.30 |\n\n## Appendix J Why Occupation Rankings Differ\n\nAcross the evaluated tasks, let\nai|c=wo⁡(i)|c/no⁡(i)a\\_{i\\mid c}=w\\_{o(i)\\mid c}/n\\_{o(i)} be task ii’s normalized\nGDP-based weight within category cc, and let Yi=1Y\\_{i}=1 when the AI output\nwins or ties and zero otherwise. Let EiE\\_{i} and CiC\\_{i} denote measured\nenergy and AI-system execution cost. The category win–tie rate and its\nenergy-normalized version are\n\n|  |  |  |  |  |\n| --- | --- | --- | --- | --- |\n|  | pc\\displaystyle p\\_{c} | =∑i∈cai|c​Yi,E¯c=∑i∈cai|c​Ei,\\displaystyle=\\sum\\_{i\\in c}a\\_{i\\mid c}Y\\_{i},\\qquad\\bar{E}\\_{c}=\\sum\\_{i\\in c}a\\_{i\\mid c}E\\_{i}, |  | (9) |\n|  |  |  |  |  |\n| --- | --- | --- | --- | --- |\n|  | WTc\\displaystyle\\mathrm{WT}\\_{c} | =100pc(%),WT/MJc=100​pcE¯c.\\displaystyle=100p\\_{c}\\ (\\%),\\qquad\\mathrm{WT/MJ}\\_{c}=\\frac{100p\\_{c}}{\\bar{E}\\_{c}}. |  | (10) |\n\nThe latter unit is percentage points per MJ. All attempts contribute energy\nand cost, including losses.\n\nFor pc>0p\\_{c}>0, let G¯S,c\\bar{G}\\_{S,c} be output dollar value per win or tie and\nD¯S,c\\bar{D}\\_{S,c} be time saving per win or tie. With\nC¯c=∑iai|c​Ci\\bar{C}\\_{c}=\\sum\\_{i}a\\_{i\\mid c}C\\_{i}, the two economic returns decompose as\n\n|  |  |  |  |\n| --- | --- | --- | --- |\n|  | NVJc=pcE¯c​G¯S,c−C¯cE¯c,TSJc=pcE¯c​D¯S,c.\\mathrm{NVJ}\\_{c}=\\frac{p\\_{c}}{\\bar{E}\\_{c}}\\bar{G}\\_{S,c}-\\frac{\\bar{C}\\_{c}}{\\bar{E}\\_{c}},\\qquad\\mathrm{TSJ}\\_{c}=\\frac{p\\_{c}}{\\bar{E}\\_{c}}\\bar{D}\\_{S,c}. |  | (11) |\n\nThus win–tie frequency per joule, dollar value per win or tie, and time\nsaving per win or tie play different roles. The category values\nare in Tables [8](#A9.T8 \"Table 8 ‣ Appendix I Occupation-Category Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\"),\n[10](#A9.T10 \"Table 10 ‣ Appendix I Occupation-Category Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\"), and [11](#A9.T11 \"Table 11 ‣ Appendix I Occupation-Category Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\").\nTable [12](#A10.T12 \"Table 12 ‣ Appendix J Why Occupation Rankings Differ ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\") shows the resulting rank changes.\nProtective Service ranks first on win–tie rate per MJ but fourth on NVJ and\nseventh on TSJ. Computer & Mathematical ranks third on win–tie rate per MJ\nbut first on NVJ because its dollar value per win or tie is higher. Production ranks\neleventh on win–tie rate per MJ and sixth on TSJ, despite the long expert\ncompletion times of its tested tasks. Architecture & Engineering leads TSJ.\nEach metric therefore answers a\ndifferent question about the same attempts.\n\nTable 12: Different categories lead win–tie rate per MJ, NVJ, and TSJ. Point-estimate ranks for GLM 5.2 against the expert-only reference, with one attempt per task and category-normalized GDP-based weights; rank 1 is highest.\n\n| Category | Win–tie pp/MJ rank | NVJ rank | TSJ rank |\n| --- | --- | --- | --- |\n| Architecture & Eng. | 8 | 2 | 1 |\n| Arts & Media | 10 | 10 | 9 |\n| Business & Financial | 9 | 8 | 8 |\n| Community & Social | 5 | 6 | 5 |\n| Computer & Math | 3 | 1 | 2 |\n| Healthcare | 12 | 11 | 12 |\n| Legal | 6 | 7 | 13 |\n| Management | 4 | 3 | 4 |\n| Office & Admin. | 13 | 13 | 11 |\n| Personal Care | 2 | 9 | 3 |\n| Production | 11 | 5 | 6 |\n| Protective Service | 1 | 4 | 7 |\n| Sales | 7 | 12 | 10 |\n\n### J.1 Full Occupation-Category Results\n\nTable [13](#A10.T13 \"Table 13 ‣ J.1 Full Occupation-Category Results ‣ Appendix J Why Occupation Rankings Differ ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\") gives the\ncategory-normalized credited value, AI execution cost, time saving,\nand energy that form each reported ratio. Table [14](#A10.T14 \"Table 14 ‣ J.1 Full Occupation-Category Results ‣ Appendix J Why Occupation Rankings Differ ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\")\ngives net value per dollar of AI-system execution cost and per AI hour for\nthe same categories. Table [15](#A10.T15 \"Table 15 ‣ J.1 Full Occupation-Category Results ‣ Appendix J Why Occupation Rankings Differ ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\") gives the\ncorresponding point estimates for every evaluated occupation.\nGDP-weighting the underlying unrounded components before forming the ratios reproduces the\nGLM 5.2 row of Table [1](#S3.T1 \"Table 1 ‣ 3.2 How do models trade off economic value and energy? (Q2) ‣ 3 Experiments and Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\") (37.69 per AI dollar\nand 191.3 USD per AI hour).\n\nTable 13: A high win–tie rate alone does not determine credited value or time saving. Category-normalized GDP-based task averages, expert-only reference. Credited value and time saving are zero on losses; AI-system execution cost and energy include every attempt. Applying (G−C)/E(G-C)/E and S/ES/E to the underlying unrounded components reproduces the NVJ and TSJ point estimates in Table [10](#A9.T10 \"Table 10 ‣ Appendix I Occupation-Category Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\").\n\n| Category | Win–tie (%) | Credited value GG (USD) | AI cost CC (USD) | Time saving SS (hours) | Energy EE (MJ) | AI time LL (hours) |\n| --- | --- | --- | --- | --- | --- | --- |\n| Architecture & Eng. | 69.0 | 206.94 | 2.104 | 3.905 | 0.495 | 0.467 |\n| Arts & Media | 46.1 | 48.70 | 1.986 | 1.219 | 0.450 | 0.384 |\n| Business & Financial | 56.8 | 75.74 | 2.172 | 1.677 | 0.505 | 0.443 |\n| Community & Social | 60.0 | 54.35 | 1.233 | 1.735 | 0.282 | 0.321 |\n| Computer & Math | 80.0 | 187.72 | 1.562 | 2.589 | 0.355 | 0.420 |\n| Healthcare | 28.2 | 45.81 | 2.306 | 0.742 | 0.445 | 0.299 |\n| Legal | 60.0 | 56.97 | 1.577 | 0.555 | 0.370 | 0.368 |\n| Management | 78.7 | 136.53 | 1.576 | 2.431 | 0.360 | 0.319 |\n| Office & Admin. | 38.1 | 26.91 | 2.727 | 1.111 | 0.604 | 0.449 |\n| Personal Care | 95.8 | 55.81 | 1.890 | 2.870 | 0.401 | 0.393 |\n| Production | 40.0 | 123.84 | 2.333 | 3.361 | 0.550 | 0.610 |\n| Protective Service | 60.0 | 63.47 | 0.976 | 1.114 | 0.225 | 0.277 |\n| Sales | 55.3 | 24.88 | 1.587 | 0.717 | 0.356 | 0.326 |\n\nTable 14: Computer & Mathematical leads net value per AI dollar and per AI hour across occupation categories. GLM 5.2 against the expert-only reference, one attempt per task. Category-normalized GDP-based averages are combined before forming (G−C)/C(G-C)/C and (G−C)/L(G-C)/L, where GG is credited output dollar value, CC is AI-system execution cost, and LL is AI-system completion time in hours. All attempts contribute cost and time. Point estimates use the underlying unrounded components summarized in Table [13](#A10.T13 \"Table 13 ‣ J.1 Full Occupation-Category Results ‣ Appendix J Why Occupation Rankings Differ ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\"); categories are ordered by net value per AI dollar.\n\n| Rank | Category | Net value/AI dollar (USD/USD) | Net value/AI hour (USD/hour) |\n| --- | --- | --- | --- |\n| 1 | Computer & Math | 119.2 | 442.9 |\n| 2 | Architecture & Eng. | 97.4 | 438.4 |\n| 3 | Management | 85.6 | 423.6 |\n| 4 | Protective Service | 64.0 | 225.8 |\n| 5 | Production | 52.1 | 199.3 |\n| 6 | Community & Social | 43.1 | 165.7 |\n| 7 | Legal | 35.1 | 150.5 |\n| 8 | Business & Financial | 33.9 | 166.1 |\n| 9 | Personal Care | 28.5 | 137.1 |\n| 10 | Arts & Media | 23.5 | 121.8 |\n| 11 | Healthcare | 18.9 | 145.3 |\n| 12 | Sales | 14.7 | 71.5 |\n| 13 | Office & Admin. | 8.9 | 53.9 |\n\nTable 15: Occupation-level economic returns vary widely across the evaluated work. Expert-only reference and one evaluated attempt per task. Win–tie is the fraction of an occupation’s evaluated tasks judged a win or tie. Each occupation’s task components are averaged before ratios are formed. These are point estimates; all attempts contribute cost and energy.\n\n| Occupation | Tasks | Win–tie (%) | NVJ (USD/MJ) | TSJ (h/MJ) |\n| --- | --- | --- | --- | --- |\n| Accountants and Auditors | 5 | 40.0 | 31.5 | 0.53 |\n| Administrative Services Managers | 5 | 100.0 | 369.8 | 6.60 |\n| Audio and Video Technicians | 5 | 40.0 | 22.7 | 0.53 |\n| Buyers and Purchasing Agents | 5 | 40.0 | 249.3 | 6.69 |\n| Child, Family, and School Social Workers | 5 | 60.0 | 188.5 | 6.16 |\n| Compliance Officers | 5 | 80.0 | 333.9 | 8.45 |\n| Computer and Information Systems Managers | 5 | 60.0 | 515.3 | 5.82 |\n| Concierges | 5 | 80.0 | 214.2 | 11.18 |\n| Counter and Rental Clerks | 5 | 60.0 | 56.4 | 2.71 |\n| Customer Service Representatives | 5 | 40.0 | 45.0 | 2.09 |\n| Editors | 5 | 80.0 | 407.1 | 10.49 |\n| Film and Video Editors | 5 | 60.0 | 116.7 | 2.94 |\n| Financial Managers | 5 | 40.0 | 40.9 | 0.33 |\n| Financial and Investment Analysts | 5 | 40.0 | 231.6 | 4.35 |\n| First-Line Supervisors of Non-Retail Sales Workers | 5 | 20.0 | 82.8 | 1.83 |\n| First-Line Supervisors of Office and Administrative Support Workers | 5 | 20.0 | 6.1 | 0.18 |\n| First-Line Supervisors of Police and Detectives | 5 | 60.0 | 292.0 | 5.08 |\n| First-Line Supervisors of Production and Operating Workers | 5 | 40.0 | 221.1 | 6.12 |\n| First-Line Supervisors of Retail Sales Workers | 5 | 60.0 | 65.2 | 2.54 |\n| General and Operations Managers | 5 | 80.0 | 1270.8 | 24.74 |\n| Industrial Engineers | 5 | 60.0 | 141.8 | 2.44 |\n| Lawyers | 5 | 60.0 | 149.7 | 1.50 |\n| Mechanical Engineers | 5 | 80.0 | 913.4 | 17.90 |\n| Medical Secretaries and Administrative Assistants | 5 | 60.0 | 90.5 | 3.70 |\n| Medical and Health Services Managers | 5 | 100.0 | 241.0 | 3.78 |\n| News Analysts, Reporters, and Journalists | 5 | 60.0 | 231.4 | 7.59 |\n| Nurse Practitioners | 5 | 80.0 | 653.5 | 9.79 |\n| Order Clerks | 5 | 20.0 | 16.0 | 0.84 |\n| Personal Financial Advisors | 5 | 60.0 | 190.4 | 3.29 |\n| Pharmacists | 5 | 60.0 | 162.8 | 2.42 |\n| Private Detectives and Investigators | 5 | 60.0 | 93.6 | 3.36 |\n| Producers and Directors | 5 | 20.0 | 39.6 | 0.91 |\n| Project Management Specialists | 5 | 60.0 | 197.1 | 3.52 |\n| Property, Real Estate, and Community Association Managers | 5 | 80.0 | 119.1 | 2.96 |\n| Real Estate Brokers | 5 | 40.0 | 26.9 | 0.58 |\n| Real Estate Sales Agents | 5 | 40.0 | 28.1 | 0.73 |\n| Recreation Workers | 5 | 100.0 | 125.4 | 6.70 |\n| Registered Nurses | 5 | 20.0 | 53.4 | 1.05 |\n| Sales Managers | 5 | 80.0 | 331.5 | 4.36 |\n| Sales Representatives, Wholesale and Manufacturing, Except Technical and Scientific Products | 5 | 60.0 | 72.4 | 1.81 |\n| Sales Representatives, Wholesale and Manufacturing, Technical and Scientific Products | 5 | 40.0 | 99.3 | 1.85 |\n| Securities, Commodities, and Financial Services Sales Agents | 5 | 100.0 | 154.3 | 3.56 |\n| Shipping, Receiving, and Inventory Clerks | 5 | 40.0 | 61.0 | 2.96 |\n| Software Developers | 5 | 80.0 | 523.8 | 7.29 |\n\n### J.2 Task-composition sensitivity\n\nThe category intervals in\nTables [8](#A9.T8 \"Table 8 ‣ Appendix I Occupation-Category Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\") and [10](#A9.T10 \"Table 10 ‣ Appendix I Occupation-Category Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\")\ncome from 10,000 task-composition bootstrap draws. Each draw resamples evaluated\ntasks within occupations, then recomputes the weighted dollar-value,\nsaved-time, cost, and energy components before forming each ratio.\nRecorded judgments and resource measurements stay fixed.\n\nWe also test category-rank stability by independently removing one uniformly\nchosen task from each occupation category in 10,000 portfolios. For each\nportfolio, we re-average the retained tasks within their occupations and\nrecompute win–tie rate per MJ, NVJ, and TSJ using the same GDP occupation\nweights. This changes task composition while keeping the recorded judgments,\nAI-system execution costs, and energy measurements fixed.\nTable [16](#A10.T16 \"Table 16 ‣ J.2 Task-composition sensitivity ‣ Appendix J Why Occupation Rankings Differ ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\") reports the resulting rank correlations and\nhow often each original leader stays first. Computer & Mathematical retains\nthe NVJ lead in 76.5% of portfolios, while Architecture & Engineering\nretains the TSJ lead in 40.4%. These frequencies describe the specified task removals, not the\nprobability that a category would lead across all professional work.\n\nTable 16: The NVJ leader survives one-task deletions more often than the TSJ leader. Across 10,000 independently drawn portfolios, one task is chosen uniformly from each category and the remaining tasks are re-averaged within their occupations. GDP occupation weights and recorded outcomes and resource use stay fixed. Percentile ranges describe these perturbations, not confidence intervals.\n\n| Ranking | Median ρ\\rho | 5th–95th ρ\\rho | Leader retained |\n| --- | --- | --- | --- |\n| Win–tie rate per MJ | 0.951 | 0.896–0.989 | 51.0% |\n| NVJ | 0.962 | 0.824–0.995 | 76.5% |\n| TSJ | 0.918 | 0.813–0.973 | 40.4% |\n\n## Appendix K Expert-Reference and Repeated-Run Sensitivity\n\n### K.1 Separating reference judgments from time valuation\n\nThe expert reference affects the metrics through its deliverable, which determines\nwhether the autonomous output wins or ties, and its expert completion time, which determines\nthe output dollar value and expert time assigned to that outcome. We cross these two choices for the same\nGLM 5.2 outputs under the one-run policy, keeping wages, occupation weights,\nAI-system completion time, AI-system execution cost, and energy fixed. The\ncalculation reuses the recorded runs and judge decisions.\n\nLet s,t∈{E,E+AI}s,t\\in\\{\\mathrm{E},\\mathrm{E{+}AI}\\} index the judgment and time\nreferences separately, with E denoting the expert-only reference and E+AI denoting the\nexpert-with-AI reference. Let YisY\\_{i}^{s} indicate a win or tie against deliverable\nss, HitH\\_{i}^{t} be expert completion time, WiW\\_{i} the median occupation\nwage, LiL\\_{i} the evaluated run’s AI-system completion time, and Ci,EiC\\_{i},E\\_{i} its AI-system execution cost and\nenergy. With normalized task weights pi=wo⁡(i)GDP/no⁡(i)p\\_{i}=w\\_{o(i)}^{\\mathrm{GDP}}/n\\_{o(i)},\nwhere no⁡(i)n\\_{o(i)} is the number of evaluated tasks in occupation o⁡(i)o(i),\n\n|  |  |  |  |  |\n| --- | --- | --- | --- | --- |\n|  | NVJ⁡(s,t)\\displaystyle\\mathrm{NVJ}(s,t) | =∑ipi​Yis​Hit​Wi−∑ipi​Ci∑ipi​Ei,\\displaystyle=\\frac{\\sum\\_{i}p\\_{i}Y\\_{i}^{s}H\\_{i}^{t}W\\_{i}-\\sum\\_{i}p\\_{i}C\\_{i}}{\\sum\\_{i}p\\_{i}E\\_{i}}, |  | (12) |\n|  |  |  |  |  |\n| --- | --- | --- | --- | --- |\n|  | TSJ⁡(s,t)\\displaystyle\\mathrm{TSJ}(s,t) | =∑ipi​Yis​(Hit−Li)∑ipi​Ei.\\displaystyle=\\frac{\\sum\\_{i}p\\_{i}Y\\_{i}^{s}(H\\_{i}^{t}-L\\_{i})}{\\sum\\_{i}p\\_{i}E\\_{i}}. |  | (13) |\n\nNegative time saving is retained. All four combinations charge\n0.429915 MJ and $1.91739 in the GDP-weighted task average.\nTable [17](#A11.T17 \"Table 17 ‣ K.1 Separating reference judgments from time valuation ‣ Appendix K Expert-Reference and Repeated-Run Sensitivity ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\") reports their returns.\n\nTable 17: Expert-with-AI judgments and completion times each lower reported NVJ and TSJ.\nFor GLM 5.2 under pass​@​1\\mathrm{pass@}1, the mixed-reference rows separate judgment and valuation effects.\nBrackets show percentile 95% task-composition intervals.\n\n| Judgment reference | Time reference | NVJ (USD/MJ) | TSJ (hours/MJ) |\n| --- | --- | --- | --- |\n| Expert-only | Expert-only | 168.11 [126.00, 222.43] | 3.823 [2.870, 5.021] |\n| Expert-with-AI | Expert-only | 69.07 [46.04, 95.84] | 1.818 [1.256, 2.486] |\n| Expert-only | Expert-with-AI | 84.90 [64.79, 107.49] | 1.990 [1.432, 2.639] |\n| Expert-with-AI | Expert-with-AI | 50.04 [30.77, 70.41] | 1.340 [0.814, 1.888] |\n\n##### Paired contrasts and uncertainty.\n\nWe resample evaluated tasks with replacement within each of the 44 occupations\n10,000 times (seed 20260726), sharing each draw across all four combinations\nand recomputing ratios and percentage reductions. In a separate sensitivity\nanalysis, we ran each task four times and compared returns using one through\nfour attempts (Appendix [K.2](#A11.SS2 \"K.2 Repeated runs: Pass@𝑘 and four-run win–tie thresholds ‣ Appendix K Expert-Reference and Repeated-Run Sensitivity ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\")). Table [18](#A11.T18 \"Table 18 ‣ Paired contrasts and uncertainty. ‣ K.1 Separating reference judgments from time valuation ‣ Appendix K Expert-Reference and Repeated-Run Sensitivity ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\")\nuses the relevant starting cell as the percentage denominator. The intervals\nvary task composition while holding recorded judgments, valuation inputs,\nand energy allocation fixed.\n\nTable 18: Changing both judgments and times reduces NVJ by 70.2% and TSJ by 65.0%.\nEvery change substitutes the expert-with-AI reference for the expert-only reference. Values are percentage\nreductions [95% interval]; a negative endpoint permits an increase.\n\n| Change and fixed quantity | NVJ reduction (%) | TSJ reduction (%) |\n| --- | --- | --- |\n| Judgments; expert-only reference times fixed | 58.9 [43.0, 72.7] | 52.5 [37.0, 66.5] |\n| Times; expert-only reference decisions fixed | 49.5 [33.1, 63.0] | 47.9 [29.1, 63.2] |\n| Times; expert-with-AI reference decisions fixed | 27.5 [-1.7, 50.8] | 26.3 [-6.2, 51.8] |\n| Judgments; expert-with-AI reference times fixed | 41.1 [20.9, 60.4] | 32.7 [7.2, 54.9] |\n| Both judgments and times | 70.2 [54.4, 82.5] | 65.0 [46.3, 79.4] |\n\n##### Interaction and attribution.\n\nThe two reductions from the expert-only reference starting point are not additive.\nSwitching deliverables first gives NVJ 168.11→69.07→50.04168.11\\to 69.07\\to 50.04. The first step removes 58.92 percentage points of the original NVJ and the subsequent time substitution removes 11.32 points, totaling 70.23%. The reverse order gives 168.11→84.90→50.04168.11\\to 84.90\\to 50.04. Averaging the two switching orders assigns 56.7% of the total NVJ decline to reference judgments and 43.3% to time valuation (TSJ: 53.5% and 46.5%).\n\n##### Judged win–tie outcomes and reference interpretation.\n\nThe AI system wins or ties less often against the expert-with-AI reference\nthan against the expert-only reference. The assisted deliverables are\ntherefore harder to match on average under this judge, and the two\nreferences can differ in which outputs win or tie. The judgment contrast\njointly captures reference content and judge behavior.\n\n### K.2 Repeated runs: Pass@kk and four-run win–tie thresholds\n\nTo test how extra attempts affect return, we execute the AI system\nindependently four times per task. Under\npass​@​k\\mathrm{pass@}k, a task earns credit if at least one of kk independent\nattempts wins or ties against the matched expert deliverable.\nThus, pass​@​1\\mathrm{pass@}1 evaluates one attempt per task, whereas\npass​@​4\\mathrm{pass@}4 asks whether any of four attempts wins or ties.\nIncreasing kk adds both another opportunity to produce a win or tie and that run’s\nenergy to the denominator.\n\nThe *four-run win–tie threshold* is a separate analysis. It always\nevaluates and charges all four runs, but varies how many of those four must win\nor tie. Pass@kk therefore changes the number of opportunities and the energy\ncounted; the four-run win–tie threshold holds both fixed and changes only the\nnumber of wins or ties required.\n\nAdditional runs improve win–tie coverage but reduce return per\njoule. Figure [10](#A11.F10 \"Figure 10 ‣ K.2 Repeated runs: Pass@𝑘 and four-run win–tie thresholds ‣ Appendix K Expert-Reference and Repeated-Run Sensitivity ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\") shows pass​@​1\\mathrm{pass@}1,\npass​@​2\\mathrm{pass@}2, and pass​@​4\\mathrm{pass@}4; NVJ and\nTSJ decline with kk because the increase in tasks with a win or tie does not offset the\nenergy of every additional run. These comparisons charge all scheduled attempts,\nincluding attempts after an earlier output wins or ties; they do not evaluate early\nstopping or show that every retry policy reduces return per joule. Retry\npolicy is therefore an efficiency decision as well as a reliability decision.\n\nFigure 10: More runs and higher four-run win–tie thresholds both lower measured NVJ and TSJ, but for different reasons. GLM 5.2 uses the expert-only reference. Left: pass@1, pass@2, and pass@4 include the energy of one, two, and four runs per task. Right: the end-to-end energy of four runs is fixed while the four-run win–tie threshold increases. Whiskers are percentile 95% task-composition bootstrap intervals.\n\nOn tasks that match the expert-with-AI reference, the AI system accounts\nfor most of the measured time saving. At the ≥1/4\\geq 1/4 threshold, at least\none of four AI runs wins or ties against the expert-with-AI deliverable. For\nthose tasks, the AI system saves 1.24 GDP-weighted hours per task relative to\nthe expert-only reference. Of this total, 0.43 hours (35.0%) is the difference\nbetween the two expert completion times, and 0.80 hours (65.0%) is the\ndifference between expert-with-AI and AI-system completion times.\nAppendix [K.1](#A11.SS1 \"K.1 Separating reference judgments from time valuation ‣ Appendix K Expert-Reference and Repeated-Run Sensitivity ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\") separates the effects of changing\nthe judged deliverable and completion time while holding AI runs and energy fixed.\n\nFigure 11: Automation adds time saving beyond expert assistance, but the balance shifts with the success threshold. At lower thresholds, most saved hours come from automation; at stricter thresholds, more come from expert assistance. GLM 5.2 tasks meet each stated win–tie threshold against the expert-with-AI reference. The left panel reports GDP-weighted hours; the right divides each transition’s saved hours by completion time at its starting expert reference. All four runs’ energy is retained. Percentages across transitions use different denominators and should not be added.\n\nFor each task and expert reference, the run order is fixed before evaluating\npass​@​k\\mathrm{pass@}k. Let Yi​rbY\\_{ir}^{b} indicate a win or tie on run rr,\nRi,kb=∑r=1kYi​rbR\\_{i,k}^{b}=\\sum\\_{r=1}^{k}Y\\_{ir}^{b} count runs that win or tie among the first\nkk, and Qi,kb=𝟏​(Ri,kb≥1)Q\\_{i,k}^{b}=\\mathbf{1}(R\\_{i,k}^{b}\\geq 1) indicate whether a task\nhas at least one win or tie. For occupation oo, the exact outcome accounting is\n\n|  |  |  |  |  |\n| --- | --- | --- | --- | --- |\n|  | Go,kb\\displaystyle G\\_{o,k}^{b} | =∑i:o⁡(i)=oQi,kbGib,\\displaystyle=\\sum\\_{i:o(i)=o}Q\\_{i,k}^{b}G\\_{i}^{b}, |  | (14) |\n|  | So,kb\\displaystyle S\\_{o,k}^{b} | =∑i:o⁡(i)=oQi,kb=1(Hib−∑r=1kYi​rb​Li​rRi,kb),\\displaystyle=\\sum\\_{\\begin{subarray}{c}i:o(i)=o\\\\ Q\\_{i,k}^{b}=1\\end{subarray}}\\left(H\\_{i}^{b}-\\frac{\\sum\\_{r=1}^{k}Y\\_{ir}^{b}L\\_{ir}}{R\\_{i,k}^{b}}\\right), |  |\n|  | Eo,k\\displaystyle E\\_{o,k} | =∑i:o⁡(i)=o∑r=1kEi​r,\\displaystyle=\\sum\\_{i:o(i)=o}\\sum\\_{r=1}^{k}E\\_{ir}, |  |\n|  | Co,k\\displaystyle C\\_{o,k} | =∑i:o⁡(i)=o∑r=1kCi​r.\\displaystyle=\\sum\\_{i:o(i)=o}\\sum\\_{r=1}^{k}C\\_{ir}. |  |\n\nOutput dollar value is counted at most once. Time saving equals expert completion time\nminus mean AI-system completion time among wins and ties when Qi,kb=1Q\\_{i,k}^{b}=1, and zero otherwise; AI-system\ncompletion times for losses are excluded. Energy and AI-system execution cost include all first kk runs without early\nstopping, regardless of their judgments. This\nbest-of-kk convention is also used by\nAPEX-Agents ([Vidgen et al., 2026](#bib.bib4)).\n\nTable [19](#A11.T19 \"Table 19 ‣ K.2 Repeated runs: Pass@𝑘 and four-run win–tie thresholds ‣ Appendix K Expert-Reference and Repeated-Run Sensitivity ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\") reports all four pass​@​k\\mathrm{pass@}k policies.\n\nTable 19: Four attempts increase output dollar value but reduce NVJ and TSJ. GLM 5.2 uses the expert-only reference. Every included run incurs cost and energy; output dollar value is counted once. NVJ is USD/MJ and TSJ is hours/MJ. Brackets give conditional 95% task-composition intervals.\n\n| Policy | Output value (USD) | AI cost (USD) | Energy (MJ) | NVJ [95% CI] | TSJ [95% CI] | Net value/ AI cost |\n| --- | --- | --- | --- | --- | --- | --- |\n| pass​@​1\\mathrm{pass@}1 | 74.19 | 1.92 | 0.430 | 168.1 [126.0, 222.4] | 3.82 [2.87, 5.02] | 37.69 |\n| pass​@​2\\mathrm{pass@}2 | 85.75 | 4.00 | 0.891 | 91.8 [71.0, 118.6] | 2.13 [1.65, 2.72] | 20.45 |\n| pass​@​3\\mathrm{pass@}3 | 89.41 | 5.98 | 1.332 | 62.7 [48.9, 80.0] | 1.46 [1.16, 1.85] | 13.96 |\n| pass​@​4\\mathrm{pass@}4 | 93.42 | 7.92 | 1.760 | 48.6 [38.3, 61.7] | 1.16 [0.93, 1.45] | 10.79 |\n\nTables [20](#A11.T20 \"Table 20 ‣ K.2 Repeated runs: Pass@𝑘 and four-run win–tie thresholds ‣ Appendix K Expert-Reference and Repeated-Run Sensitivity ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\") and [21](#A11.T21 \"Table 21 ‣ K.2 Repeated runs: Pass@𝑘 and four-run win–tie thresholds ‣ Appendix K Expert-Reference and Repeated-Run Sensitivity ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\") give complete\nGDP-weighted benchmark outcomes across four-run win–tie thresholds.\n\nTable 20: Stricter four-run thresholds reduce credited output dollar value and returns at fixed energy. GLM 5.2 uses the expert-only reference. Every included run incurs cost and energy; output dollar value is counted once. NVJ is USD/MJ and TSJ is hours/MJ. Brackets give conditional 95% task-composition intervals.\n\n| Policy | Output value (USD) | AI cost (USD) | Energy (MJ) | NVJ [95% CI] | TSJ [95% CI] | Net value/ AI cost |\n| --- | --- | --- | --- | --- | --- | --- |\n| ≥1/4\\geq 1/4 | 93.42 | 7.92 | 1.760 | 48.6 [38.3, 61.7] | 1.16 [0.93, 1.45] | 10.79 |\n| ≥2/4\\geq 2/4 | 81.58 | 7.92 | 1.760 | 41.9 [31.7, 54.9] | 1.02 [0.79, 1.31] | 9.30 |\n| ≥3/4\\geq 3/4 | 71.53 | 7.92 | 1.760 | 36.1 [26.3, 48.8] | 0.89 [0.67, 1.17] | 8.03 |\n| 4/44/4 | 57.12 | 7.92 | 1.760 | 28.0 [18.4, 40.5] | 0.72 [0.50, 0.99] | 6.21 |\n\nAgainst the expert-with-AI reference, requiring a win or tie on at least\nthree of four runs brings output dollar value below AI-system execution cost,\nso NVJ becomes negative at the two strictest thresholds.\n\nFor tasks meeting the same threshold against the expert-with-AI reference, total\nexpert-only reference-to-AI-system time saving decomposes exactly into\nexpert-only reference-to-expert-with-AI reference and\nexpert-with-AI reference-to-AI-system components.\n\nTable 21: Stricter four-run thresholds turn NVJ negative against the expert-with-AI reference. GLM 5.2 incurs cost and energy for every included run; output dollar value is counted once. NVJ is USD/MJ and TSJ is hours/MJ. Brackets give conditional 95% task-composition intervals.\n\n| Policy | Output value (USD) | AI cost (USD) | Energy (MJ) | NVJ [95% CI] | TSJ [95% CI] | Net value/ AI cost |\n| --- | --- | --- | --- | --- | --- | --- |\n| ≥1/4\\geq 1/4 | 35.11 | 7.92 | 1.760 | 15.4 [10.5, 20.7] | 0.46 [0.32, 0.61] | 3.43 |\n| ≥2/4\\geq 2/4 | 27.70 | 7.92 | 1.760 | 11.2 [6.1, 16.7] | 0.37 [0.23, 0.52] | 2.50 |\n| ≥3/4\\geq 3/4 | 7.71 | 7.92 | 1.760 | -0.1 [-1.9, 2.1] | 0.10 [0.05, 0.16] | -0.03 |\n| 4/44/4 | 3.48 | 7.92 | 1.760 | -2.5 [-3.7, -1.1] | 0.05 [0.01, 0.09] | -0.56 |\n\nTable 22: Most measured time saving occurs between the assisted expert reference and autonomous AI output. GDP-weighted task-average decomposition under pass​@​1\\mathrm{pass@}1. Bracketed values are percentile 95% task-composition bootstrap intervals.\n\n| Transition | Time saving (hours) | Time saving (%) | Energy (MJ) | TSJ (hours/MJ) |\n| --- | --- | --- | --- | --- |\n| Expert-only →\\rightarrow expert-with-AI | 0.21 [-0.04, 0.45] | 23.0 [-5.3, 45.9] | – | – |\n| Expert-with-AI →\\rightarrow AI system | 0.58 [0.34, 0.83] | 83.7 [77.8, 87.1] | 0.430 | 1.340 [0.814, 1.888] |\n| Expert-only →\\rightarrow AI system | 0.78 [0.55, 1.03] | 87.4 [84.4, 89.8] | 0.430 | 1.818 [1.256, 2.486] |\n\nFigure [11](#A11.F11 \"Figure 11 ‣ K.2 Repeated runs: Pass@𝑘 and four-run win–tie thresholds ‣ Appendix K Expert-Reference and Repeated-Run Sensitivity ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\") shows the corresponding four-run threshold\ncomparison; Table [22](#A11.T22 \"Table 22 ‣ K.2 Repeated runs: Pass@𝑘 and four-run win–tie thresholds ‣ Appendix K Expert-Reference and Repeated-Run Sensitivity ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\") gives the pass@1 decomposition.\nThe assisted-expert-to-AI-system component contributes 0.58 hours of the\n0.78-hour expert-only-to-AI-system time saving in that comparison.\n\n## Appendix L Five-Model Comparison\n\nTable [23](#A12.T23 \"Table 23 ‣ Appendix L Five-Model Comparison ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\") reports GDP-weighted model\noutcomes across the same evaluated tasks and task-composition bootstrap intervals.\n\nFigure 12: Greater output dollar value need not mean greater economic return per joule. The five models use GDP-weighted aggregates, the expert-only reference, and one attempt per task. NVJ accounts for execution cost as well as energy; TSJ captures expert time saving per unit of energy. Whiskers are percentile 95% task-composition bootstrap intervals.\n\nTable 23: GLM 5.2 leads output dollar value; Gemma 4 31B leads NVJ and TSJ among five models. The models share the evaluated tasks, the expert-only reference, and a one-attempt policy; values are GDP-weighted task averages. NVJ is USD/MJ, TSJ is hours/MJ, and net value per dollar of AI cost divides net dollar value by AI-system execution cost. Brackets give conditional 95% task-composition intervals. Cost inputs and allocation are in Appendix [G](#A7 \"Appendix G Calculating AI-System Execution Cost ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\").\n\n| Model | GPUs | Output value (USD) | AI cost (USD) | Energy (MJ) | NVJ [95% CI] | TSJ [95% CI] | Net value/ AI cost |\n| --- | --- | --- | --- | --- | --- | --- | --- |\n| GLM 5.2 | 8×\\timesH200 | 74.19 | 1.92 | 0.430 | 168.1 [126.0, 222.4] | 3.82 [2.87, 5.02] | 37.69 |\n| Qwen3.6-27B | 8×\\timesH100 | 67.81 | 4.12 | 0.516 | 123.4 [92.2, 163.2] | 2.43 [1.73, 3.28] | 15.46 |\n| Gemma 4 31B | 8×\\timesH100 | 41.38 | 0.86 | 0.097 | 417.5 [239.5, 642.9] | 10.86 [6.76, 15.72] | 47.38 |\n| GPT-OSS 120B | 8×\\timesH100 | 39.73 | 1.98 | 0.137 | 275.7 [152.5, 429.2] | 6.38 [3.71, 9.63] | 19.09 |\n| GPT-OSS 20B | 8×\\timesH100 | 35.88 | 1.12 | 0.103 | 336.0 [180.7, 536.4] | 8.26 [4.67, 12.59] | 31.03 |\n\nFigure [12](#A12.F12 \"Figure 12 ‣ Appendix L Five-Model Comparison ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\") summarizes output dollar value, NVJ, and\nTSJ, while Table [23](#A12.T23 \"Table 23 ‣ Appendix L Five-Model Comparison ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\") also provides the energy denominator,\nGDP-weighted task averages, and intervals.\n\nOccupation-category rankings are broadly similar across models,\nespecially for output dollar value and time saving. NVJ and TSJ rankings vary\nmore because the evaluated models have different end-to-end energy profiles. Table [24](#A12.T24 \"Table 24 ‣ Appendix L Five-Model Comparison ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\")\nreports all pairwise rank correlations.\n\nTable 24: Category output-dollar-value rankings agree across models more than energy-return rankings do on average. Spearman correlations across the 13 categories after including AI-system execution cost in NVJ.\n\nModel A\nModel B\nOutput dollar value\nTime saving\nNVJ\nTSJ\n\nGLM 5.2\nQwen3.6-27B\n0.830\n0.703\n0.780\n0.544\n\nGLM 5.2\nGemma 4 31B\n0.654\n0.912\n0.560\n0.731\n\nGLM 5.2\nGPT-OSS 120B\n0.637\n0.742\n0.478\n0.478\n\nGLM 5.2\nGPT-OSS 20B\n0.643\n0.824\n0.489\n0.626\n\nQwen3.6-27B\nGemma 4 31B\n0.824\n0.665\n0.549\n0.302\n\nQwen3.6-27B\nGPT-OSS 120B\n0.769\n0.659\n0.643\n0.692\n\nQwen3.6-27B\nGPT-OSS 20B\n0.808\n0.885\n0.560\n0.659\n\nGemma 4 31B\nGPT-OSS 120B\n0.802\n0.808\n0.863\n0.654\n\nGemma 4 31B\nGPT-OSS 20B\n0.907\n0.802\n0.742\n0.692\n\nGPT-OSS 120B\nGPT-OSS 20B\n0.863\n0.863\n0.879\n0.846\n\n## Appendix M Accelerator Comparisons\n\nTable [25](#A13.T25 \"Table 25 ‣ Appendix M Accelerator Comparisons ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\") reports the resulting GDP-weighted\noutcomes and economic return per joule.\n\nTable 25: B200 has the highest NVJ and TSJ among the four single accelerators. Qwen3.6-27B is evaluated against the expert-only reference under the paper’s GDP-based weights. Each accelerator is summarized from its own recorded outcomes and measured energy. Brackets give 95% task-composition intervals; costs and energy include every attempt.\n\nAccelerator\nNVJ (USD/MJ)\nTSJ (hours/MJ)\n\nNVIDIA B200\n535.5 [388.4, 721.6]\n11.684 [8.416, 15.620]\n\nNVIDIA H200\n391.3 [274.9, 532.4]\n7.293 [4.864, 10.188]\n\nNVIDIA H100\n402.2 [298.6, 531.7]\n7.243 [5.122, 9.934]\n\nAMD MI350X\n124.1 [81.7, 177.8]\n1.260 [0.450, 2.297]\n\nOne and eight H100s use the same evaluated tasks, expert-only reference,\npass​@​1\\mathrm{pass@}1 rule, and GDP weights, with separately recorded outcomes.\nTheir respective output dollar values, execution costs, and energies are\n($64.79,$0.71,0.159​MJ)(\\$64.79,\\$0.71,0.159\\,\\mathrm{MJ}) and\n($67.81,$4.12,0.516​MJ)(\\$67.81,\\$4.12,0.516\\,\\mathrm{MJ}), yielding NVJ of 402.16 versus\n123.39 USD/MJ and TSJ of 7.243 versus 2.432 hours/MJ.\n\n## Appendix N GDPWorks: 83 Additional Occupational Tasks\n\nTo broaden professional-work evaluation beyond GDPval, we assembled\n*GDPWorks*, a dataset of 83 additional tasks created by Snorkel AI. Each task includes\na request, grading rubric, source files, expert-only and expert-with-AI\ndeliverables, and time annotations. Together with OpenAI’s 220 public GDPval\ntasks and the annotations added by Snorkel AI, these materials are prepared\nfor release by the Scaling Intelligence Lab and HazyResearch at\n[ScalingIntelligence/joules-to-gdp](https://huggingface.co/datasets/ScalingIntelligence/joules-to-gdp),\nwith GDPWorks in the joules\\_to\\_gdp split.\n\nGDPWorks adds 36 detailed O\\*NET occupations absent from GDPval’s 44,\nbringing the combined coverage to 80 distinct occupation codes. Examples\ninclude bioinformatics scientists, environmental scientists, and purchasing\nmanagers. The tasks introduce recent work contexts and 2026-dated source\nmaterials, including financial records and environmental monitoring data.\nThey also target more demanding, source-intensive workflows: median supplied\nexpert-only time estimates are 4.0 hours, compared with 2.34 hours for GDPval,\nand the median number of included reference files is six, compared with one.\nThese time estimates characterize expected expert effort. GDPWorks\nsupports future tests of how economic return per joule generalizes to\nadditional occupations and longer workflows. GDPWorks is not a\nGDP-weighted sample; the results reported in this paper use the 220-task\nGDPval split.","body_html":"<h1 id=\"joules-to-gdp-how-much-economic-value-do-ai-systems-produce-per-\">Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?</h1>\n<p>Amirreza Zeinali  ⋅\\cdot Jon Saad-Falcon  ⋅\\cdot Avanika Narayan</p>\n<p>Mersad Abbasi  ⋅\\cdot Jonathan S. Hartley  ⋅\\cdot John Hennessy\nAffiliation: Stanford University  ⋅\\cdot University of California, Berkeley\nAffiliation: The University of Texas at Austin</p>\n<p>Christopher Re  ⋅\\cdot Azalia Mirhoseini</p>\n<p>11footnotetext: Equal contribution.</p>\n<h2 id=\"1-introduction\">1 Introduction</h2>\n<p>Existing studies assess an AI system’s economic value through proxies such as output\nquality and estimated worker time saving, and compare task performance with\nAPI cost or completion time\n(<a href=\"#bib.bib2\">Patwardhan et al., 2025</a>; <a href=\"#bib.bib30\">Wan et al., 2026</a>; <a href=\"#bib.bib14\">Erol et al., 2026</a>).\nMeanwhile, energy benchmarks measure how much energy AI systems consume, but\ndo not quantify the economic value of the professional outputs produced with\nthat energy\n(<a href=\"#bib.bib16\">Chung et al., 2025</a>; <a href=\"#bib.bib33\">Tripathy et al., 2026</a>, see Appendix <a href=\"#A1\" title=\"Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">A</a> for related work).\nWe connect these approaches by measuring the economic value of professional\noutputs alongside the energy required to produce them.</p>\n<p>For this study, we use energy as our resource measure for three reasons. (1) A joule is a stable physical\nunit, while API providers can change prices. (2) End-to-end energy combines\nmodel and hardware choice, tool use, waits, and retries in one measure.\n(3) As we show, energy reveals differences that API cost or completion time\ncan miss.</p>\n<p>Having measured the energy used to produce each AI output, we then assess its\neconomic value in two ways: (1) net dollar value and (2) expert time saving\nfrom using the AI system. We express economic return using Net Value per Joule (NVJ) in USD/MJ and\nTime Saving per Joule (TSJ) in hours/MJ. We use both because the work\nwith the greatest dollar return need not save the most expert time, as our\nresults show. Figure <a href=\"#S1.F1\" title=\"Figure 1 ‣ 1 Introduction ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">1</a> summarizes this evaluation.</p>\n<p>Figure 1: Overview of Joules-to-GDP. (a) NVJ and TSJ express dollar and expert time benefits per joule. (b) The study compares occupations, models, and accelerators. (c) Occupation returns vary widely, with different NVJ and TSJ leaders. (d) More valuable output can require more energy; the frontier uses unweighted totals on log scales. The bottom panel uses 2.52 MJ, equivalent to one H200 drawing 700 W for an hour or about 23 hours of active laptop use.</p>\n<p>We apply these metrics to 220 public GDPval benchmark tasks across 44\noccupations and 13 categories, weighted by sector GDP and employment.\nWe then investigate three questions.</p>\n<p>Q1. How does economic return per joule vary across occupations?\nAcross occupation groups, concierge and recreation work, such as preparing\ntravel itineraries, leads in success rate before accounting for energy, while\nengineering leads in net dollar value per task. Accounting for energy shifts\nthe success-rate leader to investigation work, such as turning retail\ninvestigation notes into reports, and the net-value leader to software\ndevelopment, such as building website features. Factory supervision, such as\nplanning schedules to clear backlogs, falls from second in time saving per\ntask to sixth in time saving per joule. These returns depend jointly on\nsuccessful completion, the net dollar value or time saving each output\nrepresents, and energy consumption. The expert reference used to assess AI\noutputs also matters. Comparing against an expert who used AI changes the\nmeasured returns (Section <a href=\"#S3.SS1\" title=\"3.1 How does economic return per joule vary across occupations? (Q1) ‣ 3 Experiments and Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">3.1</a>).</p>\n<p>Q2. How do models trade off economic value and energy?\nAmong five evaluated models, GLM 5.2 produces the highest average output\ndollar value, at $74.19 per task, and saves 1.64 hours of expert time using\n0.430 MJ, equivalent in energy to about four hours of active laptop use. By\ncomparison, Gemma 4 31B produces more than half that output dollar value and\nsaves about two-thirds as much expert time using less than a quarter of the\nenergy, yielding approximately 2.5 times the NVJ and 2.8 times the TSJ.\nTo capture both sides of this tradeoff, we use NVJ and TSJ to compare returns\nfor the same amount of energy, alongside value–energy frontiers that show\nthe output dollar value or time saving achieved at each model’s measured\nenergy use. Both views matter because a high return per joule can coexist\nwith less valuable output per task. We also find that energy reveals\ndifferences that cost and completion time miss. GPT-OSS 120B uses 68% less\nenergy than GLM 5.2 despite nearly identical execution cost and completion\ntime (Section <a href=\"#S3.SS2\" title=\"3.2 How do models trade off economic value and energy? (Q2) ‣ 3 Experiments and Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">3.2</a>).</p>\n<p>Q3. How do accelerator type and count affect return per joule?\nAcross four tested single-accelerator deployments of Qwen3.6-27B, B200 leads\nNVJ and TSJ. We then vary accelerator count with the same Qwen3.6-27B model.\nOne and eight H100s produce nearly identical net dollar value, yet eight H100s\nreduce completion time by only 22% while using 3.24 times the energy\n(Section <a href=\"#S3.SS3\" title=\"3.3 How do accelerator type and count affect return per joule? (Q3) ‣ 3 Experiments and Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">3.3</a>).</p>\n<p>Our framework uses measured energy to assess an AI system’s economic value\nin occupational work and compare models and accelerators for deployment.\nOptimizing NVJ and TSJ targets unsuccessful attempts that consume\nenergy and additional computation that delivers little extra value, two\nshortcomings of the evaluated systems. These targets can guide improvements in\nmodels, accelerators, software, and task routing for deployments under real\nenergy constraints, including expanding on-premises and edge use\n(<a href=\"#bib.bib53\">Deloitte, 2026</a>). Our estimates use benchmark tasks, recorded expert\ntimes, and occupation median wages. Whether the same returns hold beyond this\nbenchmark remains to be tested.</p>\n<h2 id=\"2-methods\">2 Methods</h2>\n<p>We value each AI output using an expert reference, judge the output,\nmeasure energy, and calculate AI-system execution cost before aggregating across occupations. An\n<em>AI system</em> is the model, agent harness, inference backend, accelerator,\nand batching policy; a <em>run</em> is one execution on a task. The main analysis\nuses one run per task for each configuration. We judge that output,\nrecord its AI-system completion time, and charge its measured energy and\nAI-system execution cost.\nWe also collected four independent runs per task to test repeated-attempt\nsensitivity (Appendix <a href=\"#A11\" title=\"Appendix K Expert-Reference and Repeated-Run Sensitivity ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">K</a>, Figure <a href=\"#A11.F10\" title=\"Figure 10 ‣ K.2 Repeated runs: Pass@𝑘 and four-run win–tie thresholds ‣ Appendix K Expert-Reference and Repeated-Run Sensitivity ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">10</a>).</p>\n<h3 id=\"2-1-dataset-and-expert-references\">2.1 Dataset and expert references</h3>\n<p>We evaluate 220 public GDPval professional tasks spanning 44 occupations\nin nine U.S. economic sectors, with five tasks\nper occupation (<a href=\"#bib.bib2\">Patwardhan et al., 2025</a>). Each task contains a professional\nrequest, source files, grading rubric, and target deliverable. Deliverables\ninclude documents, spreadsheets, presentations, code, and media. GDPval drew\nthe tasks from professional work and applied multiple rounds of review;\nAppendix <a href=\"#A2\" title=\"Appendix B Expert References and Annotations ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">B</a> gives dataset details, and\nAppendix <a href=\"#A3\" title=\"Appendix C Benchmark Contamination ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">C</a> reports supplementary dataset checks.</p>\n<p>For each task, an expert reference consists of a finished deliverable\nand the recorded time the expert took to produce it. We use two reference\nconditions. The expert-only reference was produced by an expert\nwithout AI assistance, and the expert-with-AI reference was\nproduced by an expert using AI. Each condition has its own deliverable and\nrecorded expert completion time.\nWe use the expert-only reference as primary because its deliverable and output\ndollar value are independent of a particular assisting AI model. The\nexpert-with-AI reference provides a secondary comparison for assisted work.</p>\n<p>The expert-only record covers task setup, source review, execution, and\nchecking without AI use. The expert-with-AI record covers the expert’s\nsetup, prompting, monitoring, review, and correction, including waits for\nAI-generated output.\nAI-system completion time is elapsed time until a run stops, including losses\nand timeouts; it covers inference, tool use, and task-specific setup within\nthe request.</p>\n<p>Let GibG_{i}^{b} denote the dollar value assigned to an AI output on task ii\nif it wins or ties against expert reference bb. Let task ii belong to\noccupation o⁡(i)o(i), and let\nb∈{E,E+AI}b\\in{\\mathrm{E},\\mathrm{E{+}AI}} identify the expert reference.\nFollowing the original GDPval valuation approach (<a href=\"#bib.bib2\">Patwardhan et al., 2025</a>), we multiply expert completion time by the occupation’s median hourly wage:</p>\n<div class=\"table-wrap\"><table><thead><tr><th></th><th></th><th></th><th></th></tr></thead><tbody><tr><td></td><td>Gib=Hib​Wo⁡(i)median,G_{i}^{b}=H_{i}^{b}W_{o(i)}^{\\mathrm{median}},</td><td></td><td>(1)</td></tr></tbody></table></div>\n<p>where HibH_{i}^{b} is the recorded expert completion time for the selected reference. The term Wo⁡(i)medianW_{o(i)}^{\\mathrm{median}} is the\noccupation’s median hourly wage from the annual Occupational Employment and\nWage Statistics (OEWS) program, the U.S. national statistical reference for\noccupational employment and wage estimates. Appendix <a href=\"#A8\" title=\"Appendix H Economic Weights and Wage Robustness ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">H</a>\ndocuments the wage source and GDP-based aggregation.</p>\n<h3 id=\"2-2-blinded-evaluation\">2.2 Blinded evaluation</h3>\n<p>The AI system receives the task and source files but not its rubric.\nWe use the multimodal Gemini 3.1 Flash-Lite as the judge throughout the main\nanalysis (<a href=\"#bib.bib24\">Google DeepMind, 2026a</a>). It compares the AI-system\ndeliverable with the selected expert reference in a blinded pairwise comparison.\nThe judge receives the task, rubric,\nsource evidence, and identity-masked candidates. A shared pipeline extracts\nsanitized semantic content and rendered visual evidence, enabling the judge to\nassess the documents, spreadsheets, presentations, code, and media represented\nin GDPval. Candidate positions are balanced. The judge chooses either candidate\nor a tie. Let Yib=1Y_{i}^{b}=1 when the AI\nsystem wins or ties and Yib=0Y_{i}^{b}=0 when it loses. Following GDPval’s blinded\nwin-rate approach (<a href=\"#bib.bib2\">Patwardhan et al., 2025</a>), we measure task success by the\nwin–tie rate: the share of AI outputs judged a win or tie against the expert\nreference. The output dollar value\ncontribution is Yib​GibY_{i}^{b}G_{i}^{b}, which equals GibG_{i}^{b} for a win or tie and zero for a\nloss. We choose this binary win-or-tie/loss rule to be conservative about partial credit.\nA loss contributes zero even if the output has some utility, while a tie receives\nthe full output dollar value.\nIn a judge-consistency check, Gemini 3.1 Flash-Lite, Claude Haiku 4.5,\nand GPT-5 mini showed 72.3–77.2% pairwise exact agreement. Gemini’s\nwin–tie/loss decisions also agreed with human experts on 23 of 32 GLM 5.2\ncomparisons (71.9%).\nAppendix <a href=\"#A4\" title=\"Appendix D Judge and Evidence Pipeline ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">D</a> details the evidence pipeline, candidate balancing,\nand judge settings; Appendix <a href=\"#A4.SS1\" title=\"D.1 Judge consistency ‣ Appendix D Judge and Evidence Pipeline ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">D.1</a> reports agreement\namong the three judges and with human reviewers.\nFigure <a href=\"#S2.F2\" title=\"Figure 2 ‣ 2.2 Blinded evaluation ‣ 2 Methods ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">2</a>\nillustrates the accounting for a GDPval task.</p>\n<p>Figure 2: An illustrative GDPval example shows the evaluation workflow. An Accountants and Auditors task, Fall Music Tour P&amp;L, compares the expert-only output with the AI-agent output through a blinded LLM-as-a-Judge comparison. A win or tie earns output dollar value and time saving; every attempt incurs AI-system execution cost and energy.</p>\n<h3 id=\"2-3-end-to-end-energy-accounting\">2.3 End-to-end energy accounting</h3>\n<p>We measure energy with Intelligence per Watt (IPW), from immediately before\neach batch’s runs launch until the final run in that batch completes\n(<a href=\"#bib.bib19\">Saad-Falcon et al., 2025</a>).\nHere, <em>end-to-end</em> denotes the full execution window, including inference,\ntool use, and waits. We sum the component-energy readings recorded by IPW\n(e.g., GPU, CPU, and DRAM), using counter differences or integrated sampled power.\nAppendix <a href=\"#A5\" title=\"Appendix E Execution and End-to-End Energy Accounting ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">E</a> details the measurement procedure.</p>\n<p>We batch concurrent requests because deployed inference commonly serves\nmultiple requests at the same time. The batch policy therefore represents a\nrealistic serving condition rather than serial execution of isolated requests.</p>\n<p>For a batch of KbK_{b} requests, active-request attribution divides each\nmoment’s energy among the requests still running. For request ii,</p>\n<div class=\"table-wrap\"><table><thead><tr><th></th><th></th><th></th><th></th></tr></thead><tbody><tr><td></td><td>Ei=∫i​runningPbatch​(t)Nrunning​(t)​𝑑t+Eb,0Kb.E_{i}=\\int_{i\\ \\mathrm{running}}\\frac{P_{\\mathrm{batch}}(t)}{N_{\\mathrm{running}}(t)}\\,dt+\\frac{E_{b,0}}{K_{b}}.</td><td></td><td>(2)</td></tr></tbody></table></div>\n<p>Here Pbatch​(t)P_{\\mathrm{batch}}(t) is batch power, Nrunning​(t)N_{\\mathrm{running}}(t) counts\nactive requests, and Eb,0E_{b,0} is shared energy with no active request.\nThe active portion assigns energy according to how long each request actually\nruns, giving longer-running requests a larger share when power and concurrency\nare comparable.\nWhen only a batch total is retained, average power is total joules divided\nby batch completion time. Every request incurs energy, including losses and timeouts.\nAppendix <a href=\"#A6\" title=\"Appendix F Energy-Attribution Sensitivity ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">F</a> tests attribution sensitivity;\nFigure <a href=\"#A6.F8\" title=\"Figure 8 ‣ Appendix F Energy-Attribution Sensitivity ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">8</a> shows energy use as requests finish.</p>\n<h3 id=\"2-4-net-value-per-joule-and-time-saving-per-joule\">2.4 Net Value per Joule and Time Saving per Joule</h3>\n<p>For an AI output on task ii, let CiC_{i}, EiE_{i}, and LiL_{i} be its\nAI-system execution cost, measured energy, and AI-system completion time.\nIts credited output dollar value is Yib​GibY_{i}^{b}G_{i}^{b}, where GibG_{i}^{b} is based on\nthe selected reference expert’s completion time HibH_{i}^{b} (Equation <a href=\"#S2.E1\" title=\"In 2.1 Dataset and expert references ‣ 2 Methods ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">1</a>).\nThe output’s net dollar value is Yib​Gib−CiY_{i}^{b}G_{i}^{b}-C_{i}, and its time saving\nis Yib​(Hib−Li)Y_{i}^{b}(H_{i}^{b}-L_{i}). For occupation oo, let\nGob=∑i:o⁡(i)=oYibGibG_{o}^{b}=\\sum_{i:o(i)=o}Y_{i}^{b}G_{i}^{b}, Co=∑i:o⁡(i)=oCiC_{o}=\\sum_{i:o(i)=o}C_{i}, and\nEo=∑i:o⁡(i)=oEiE_{o}=\\sum_{i:o(i)=o}E_{i}. We define\nNet Value per Joule (NVJ) (in USD/MJ) and\nTime Saving per Joule (TSJ) (in hours/MJ):</p>\n<div class=\"table-wrap\"><table><thead><tr><th></th><th></th><th></th><th></th></tr></thead><tbody><tr><td></td><td>NVJob=Gob−CoEo,TSJob=∑i:o⁡(i)=oYib(Hib−Li)Eo.\\mathrm{NVJ}<em>{o}^{b}=\\frac{G</em>{o}^{b}-C_{o}}{E_{o}},\\qquad\\mathrm{TSJ}<em>{o}^{b}=\\frac{\\sum</em>{i:o(i)=o}Y_{i}^{b}\\left(H_{i}^{b}-L_{i}\\right)}{E_{o}}.</td><td></td><td>(3)</td></tr></tbody></table></div>\n<p>Only wins or ties contribute output dollar value or time saving; every run incurs\ncost, so a loss has negative net dollar value. We report energy in MJ\n(1​MJ=106​J1\\,\\mathrm{MJ}=10^{6}\\,\\mathrm{J}).</p>\n<p>We calculate AI-system execution cost from recorded resource\nallocations and published GPU, CPU, and RAM prices. Shared node\ntime uses active-request shares within the recorded task-execution\nwindow. Shared idle time inside that window is assigned once. Storage and\ntransfer charges are added where applicable. This method values recorded\nresource use at published rates; rate sources and allocation rules are in\nAppendix <a href=\"#A7\" title=\"Appendix G Calculating AI-System Execution Cost ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">G</a>.</p>\n<h3 id=\"2-5-two-levels-of-gdp-based-aggregation\">2.5 Two levels of GDP-based aggregation</h3>\n<p>For each evaluated occupation, we average output dollar value, AI-system execution\ncost, time saving, and energy over its observed tasks. GDP-based occupation\nweights woGDPw_{o}^{\\mathrm{GDP}} combine sector GDP shares with within-sector\nemployment shares. Wages do not enter these weights because the occupation’s median\nwage already enters output dollar value, and wage-bill weights would give\nhigh-wage occupations additional influence. These weights define a\nGDP-weighted benchmark portfolio (Appendix <a href=\"#A8\" title=\"Appendix H Economic Weights and Wage Robustness ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">H</a>, Figure <a href=\"#A8.F9\" title=\"Figure 9 ‣ Appendix H Economic Weights and Wage Robustness ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">9</a>).\nThe O*NET-SOC taxonomy groups the 44 occupations into 13 categories. We\naggregate at two levels.</p>\n<p>Occupation-category level. We renormalize occupation weights\nwithin each category to sum to one, yielding category-normalized averages\nfor comparisons among categories.</p>\n<p>Benchmark level. We weight category outcomes by their category\nweights, yielding GDP-weighted task averages for comparisons among AI systems.</p>\n<p>Writing non_{o} for the number of evaluated tasks in occupation oo,\nSob=∑i:o⁡(i)=oYib(Hib−Li)S_{o}^{b}=\\sum_{i:o(i)=o}Y_{i}^{b}(H_{i}^{b}-L_{i}),\nWc=∑o∈cwoGDPW_{c}=\\sum_{o\\in c}w_{o}^{\\mathrm{GDP}}, and X∈{G,C,S,E}X\\in{G,C,S,E}:</p>\n<div class=\"table-wrap\"><table><thead><tr><th></th><th></th><th></th><th></th></tr></thead><tbody><tr><td></td><td>X~c=∑o∈cwoGDPWc​Xono,X~GDP=∑cWc​X~c.\\widetilde{X}<em>{c}=\\sum</em>{o\\in c}\\frac{w_{o}^{\\mathrm{GDP}}}{W_{c}}\\frac{X_{o}}{n_{o}},\\qquad\\widetilde{X}^{\\mathrm{GDP}}=\\sum_{c}W_{c}\\widetilde{X}_{c}.</td><td></td><td>(4)</td></tr></tbody></table></div>\n<p>We aggregate the components first and then form the NVJ and TSJ ratios.</p>\n<h2 id=\"3-experiments-and-results\">3 Experiments and Results</h2>\n<p>The results follow Q1–Q3. Q1 holds the AI system fixed and varies the occupation\nand expert reference. Q2 compares models on the same tasks and evaluation protocol\nacross value, AI-system execution cost, AI-system completion time, and energy.\nQ3 varies accelerator type and count for the same model. This\nseparates differences in the work from differences in the AI system.</p>\n<p>Figure 3: Energy changes occupation rankings, and each return metric favors different work. GLM 5.2 against the expert-only reference, with category-normalized GDP weights. Top panels show win–tie rate, net dollar value, and time saving per task; bottom panels normalize each by energy. Categories follow the win–tie-per-MJ order; whiskers show 95% task-composition intervals.</p>\n<h3 id=\"3-1-how-does-economic-return-per-joule-vary-across-occupations-q\">3.1 How does economic return per joule vary across occupations? (Q1)</h3>\n<p>We evaluate GLM 5.2 with OpenHands on eight NVIDIA H200 GPUs in 20-request\nbatches across the evaluated tasks.</p>\n<p>Energy changes occupation rankings, and each return metric captures a different benefit.\nPersonal Care &amp; Service leads in win–tie rate, but once energy is included,\nProtective Service leads per MJ. Net dollar value shows a similar reversal.\nArchitecture &amp; Engineering (for example, CAD design) delivers about\n10% more per task than Computer &amp; Mathematical (for\nexample, website design). However, after energy is included, Computer &amp;\nMathematical delivers about 27% higher NVJ. Likewise, Production falls from second in\ntime saving per task to sixth in TSJ (Figure <a href=\"#S3.F3\" title=\"Figure 3 ‣ 3 Experiments and Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">3</a>).</p>\n<p>The three energy-normalized metrics also capture different benefits.\nFor example, Computer &amp; Mathematical delivers nearly four times Personal Care &amp;\nService’s NVJ despite a lower win–tie rate per MJ, because each successful\noutput represents more dollar value. Time saving reveals another reversal.\nProduction has about 20% lower NVJ than Protective Service, yet 23% higher\nTSJ because each success saves more expert time. High NVJ and TSJ thus combine success\nwith greater net dollar value or time saving relative to energy consumed\nacross all attempts.\nAppendices <a href=\"#A9\" title=\"Appendix I Occupation-Category Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">I</a> and <a href=\"#A10\" title=\"Appendix J Why Occupation Rankings Differ ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">J</a> give category\nvalues and explain these rankings; Appendix <a href=\"#A10.SS1\" title=\"J.1 Full Occupation-Category Results ‣ Appendix J Why Occupation Rankings Differ ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">J.1</a>\nlists full results, Appendix <a href=\"#A10.SS2\" title=\"J.2 Task-composition sensitivity ‣ Appendix J Why Occupation Rankings Differ ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">J.2</a> tests rank stability,\nand Appendix <a href=\"#A4.SS1\" title=\"D.1 Judge consistency ‣ Appendix D Judge and Evidence Pipeline ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">D.1</a> checks judge agreement.</p>\n<p>Figure 4: An expert-with-AI reference reduces GLM 5.2’s NVJ by 70% and TSJ by 65%. Only the expert deliverable and its completion time change; AI runs, costs, energy, and GDP weights stay fixed. Whiskers show 95% task-composition intervals (Appendix <a href=\"#A11.SS1\" title=\"K.1 Separating reference judgments from time valuation ‣ Appendix K Expert-Reference and Repeated-Run Sensitivity ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">K.1</a>).</p>\n<p>The expert-with-AI reference is harder to beat and closer to how many professionals work.\nThe expert-only reference provides a comparison independent of an assisting\nmodel. By contrast, the expert-with-AI reference reflects the model and tools\nthe expert used, while testing the additional benefit of automation when experts\nalready use AI. Under our weighting, experts using AI spend about 23% less\ntime on average, and GLM 5.2 matches or exceeds their deliverables about half\nas often as the expert-only deliverables. Automation must therefore meet a\nhigher quality bar while replacing less expert time, lowering NVJ and TSJ\n(Figure <a href=\"#S3.F4\" title=\"Figure 4 ‣ 3.1 How does economic return per joule vary across occupations? (Q1) ‣ 3 Experiments and Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">4</a>). Yet substantial time-saving potential\nremains. On tasks where GLM 5.2 matches or exceeds the assisted expert’s\ndeliverable, its completion time is about 84% shorter on average than the assisted\nexpert’s (Appendix <a href=\"#A11.SS2\" title=\"K.2 Repeated runs: Pass@𝑘 and four-run win–tie thresholds ‣ Appendix K Expert-Reference and Repeated-Run Sensitivity ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">K.2</a>, Table <a href=\"#A11.T22\" title=\"Table 22 ‣ K.2 Repeated runs: Pass@𝑘 and four-run win–tie thresholds ‣ Appendix K Expert-Reference and Repeated-Run Sensitivity ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">22</a>).\nFigure <a href=\"#A11.F11\" title=\"Figure 11 ‣ K.2 Repeated runs: Pass@𝑘 and four-run win–tie thresholds ‣ Appendix K Expert-Reference and Repeated-Run Sensitivity ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">11</a> compares time saving from expert assistance and\nautomation across success thresholds.</p>\n<p>Figure 5: Higher total economic benefit can come with lower return per joule. For example, GLM 5.2 delivers more total value and time saving than Gemma 4 31B, but has lower NVJ and TSJ. The top row compares GDP-weighted NVJ (a) and TSJ (b); the bottom shows unweighted total output dollar value (c) and time saving (d) versus total energy. Dashed steps connect models on each point-estimate efficient frontier. Whiskers show 95% task-composition intervals.</p>\n<h3 id=\"3-2-how-do-models-trade-off-economic-value-and-energy-q2\">3.2 How do models trade off economic value and energy? (Q2)</h3>\n<p>We serve five models locally so that we can measure end-to-end energy directly:\nGLM 5.211\n1\nGLM 5.2 uses eight H200 GPUs because its FP8 serving setup\ndid not fit on the tested eight-H100 node used for the other four models.,\nQwen3.6-27B, Gemma 4 31B, GPT-OSS 120B, and GPT-OSS 20B\n(<a href=\"#bib.bib44\">Z.ai, 2026</a>; <a href=\"#bib.bib45\">Qwen Team, 2026</a>; <a href=\"#bib.bib46\">Google DeepMind, 2026b</a>; <a href=\"#bib.bib47\">OpenAI, 2025a</a>).\nWe swap the model while holding the evaluated tasks, expert-only reference, judge,\nOpenHands harness, and vLLM inference engine fixed; the four non-GLM models\nalso use the same eight-H100 hardware.\nAppendix <a href=\"#A12\" title=\"Appendix L Five-Model Comparison ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">L</a> summarizes model outcomes (Figure <a href=\"#A12.F12\" title=\"Figure 12 ‣ Appendix L Five-Model Comparison ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">12</a>);\nAppendix <a href=\"#A6\" title=\"Appendix F Energy-Attribution Sensitivity ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">F</a> tests energy attribution for the\nGLM 5.2 occupation comparisons.</p>\n<p>Higher output need not justify its energy cost.\nGLM 5.2 produces the greatest output dollar value, $74.19 per task, and saves\n1.64 expert hours per task on average. Gemma 4 31B produces\n44% less output dollar value but uses 77% less energy,\ngiving it about 2.5 times GLM 5.2’s NVJ and the highest TSJ among these models.\nFor scale, GLM 5.2’s 0.430 MJ per task is equivalent in energy to about\nfour hours of active laptop use22\n2\nLaptop-use equivalents assume\n30 W, the midpoint of the 20–40 W active-laptop range reported by\n<a href=\"#bib.bib54\">National Renewable Energy Laboratory (2011)</a>. or 10.2 minutes of one H200 drawing 700 W\ncontinuously (<a href=\"#bib.bib20\">NVIDIA, 2026</a>).\nQwen3.6-27B produces the second-highest output dollar value but uses the most\nenergy among the five models. Figure <a href=\"#S3.F5\" title=\"Figure 5 ‣ 3.1 How does economic return per joule vary across occupations? (Q1) ‣ 3 Experiments and Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">5</a> ranks the five models by GDP-weighted\nNVJ and TSJ and shows unweighted total output dollar value and total time saving against\ntotal energy across evaluated tasks. On both unweighted point-estimate efficient\nfrontiers, Gemma 4 31B anchors the low-energy end and GLM 5.2 the high-benefit\nend. GPT-OSS 120B lies between them on the dollar-value frontier. These\ntradeoffs explain why the rankings and frontiers answer complementary questions.\nNVJ and TSJ measure net dollar return and expert time saved per megajoule,\nwhile the frontiers show how much total value or time saving each model delivers\nfor a given energy budget. A high return per joule\ncan still come from a model that delivers too little total work. These\nviews explain why Gemma 4 31B leads in return per joule while GLM 5.2 remains\non the frontier by delivering more value at higher energy use.\nOccupation-category rankings are broadly similar across models\n(Appendix <a href=\"#A12\" title=\"Appendix L Five-Model Comparison ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">L</a>).</p>\n<p>Table 1: GLM 5.2 and GPT-OSS 120B have similar cost and completion time but different energy returns.\nAI-system execution cost, AI-system completion time, and energy are GDP-weighted\ntask averages; the three efficiency\nratios share each system’s net-dollar-value numerator. Colors identify systems.\nBars show point estimates on a separate zero-to-maximum scale within each\nefficiency column; longer bars mean greater net value per unit of resource.</p>\n<div class=\"table-wrap\"><table><thead><tr><th></th><th></th><th></th><th></th><th></th><th></th><th></th></tr></thead><tbody><tr><td></td><td>GDP-weighted task averages</td><td></td><td></td><td>Net value per</td><td></td><td></td></tr><tr><td>Model</td><td>Execution cost (USD)</td><td>Completion time (min)</td><td>Energy (MJ)</td><td>dollar of AI cost (USD/USD)</td><td>AI hour (USD/hour)</td><td>megajoule (USD/MJ)</td></tr><tr><td>GLM 5.2</td><td>1.92</td><td>22.7</td><td>0.430</td><td>37.69</td><td>191.3</td><td>168.1</td></tr><tr><td>Qwen3.6-27B</td><td>4.12</td><td>51.0</td><td>0.516</td><td>15.46</td><td>75.0</td><td>123.4</td></tr><tr><td>Gemma 4 31B</td><td>0.86</td><td>7.7</td><td>0.097</td><td>47.38</td><td>316.9</td><td>417.5</td></tr><tr><td>GPT-OSS 120B</td><td>1.98</td><td>22.2</td><td>0.137</td><td>19.09</td><td>102.1</td><td>275.7</td></tr><tr><td>GPT-OSS 20B</td><td>1.12</td><td>15.3</td><td>0.103</td><td>31.03</td><td>136.3</td><td>336.0</td></tr></tbody></table></div>\n<h4 id=\"3-2-1-net-value-per-dollar-of-ai-cost-and-per-ai-hour\">3.2.1 Net value per dollar of AI cost and per AI hour</h4>\n<p>To test whether energy as a denominator reveals a dimension of efficiency\nthat monetary cost and AI-system completion time alone miss, we define two comparison metrics:</p>\n<div class=\"table-wrap\"><table><thead><tr><th></th><th></th><th></th><th></th></tr></thead><tbody><tr><td></td><td>Net value/dollar of AI cost=G−CC,Net value/AI hour=G−CL.\\text{Net value/dollar of AI cost}=\\frac{G-C}{C},\\qquad\\text{Net value/AI hour}=\\frac{G-C}{L}.</td><td></td><td>(5)</td></tr></tbody></table></div>\n<p>Here GG, CC, and LL are GDP-weighted averages of credited output dollar\nvalue, AI-system execution cost, and AI-system completion time. Costs and durations include\nlosses and timeouts; concurrent runs each contribute their AI-system completion time to LL, so an\nAI hour is a request-hour. Table <a href=\"#S3.T1\" title=\"Table 1 ‣ 3.2 How do models trade off economic value and energy? (Q2) ‣ 3 Experiments and Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">1</a> compares\nresource use and the three efficiency ratios, each with numerator G−CG-C.\nAppendix <a href=\"#A7\" title=\"Appendix G Calculating AI-System Execution Cost ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">G</a> details resource prices and cost allocation;\nAppendix <a href=\"#A7.SS1\" title=\"G.1 Execution-cost sensitivity ‣ Appendix G Calculating AI-System Execution Cost ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">G.1</a> tests alternative provider rates.</p>\n<p>Energy reveals differences beyond AI-system execution cost and completion time.\nGLM 5.2 and GPT-OSS 120B have nearly the same GDP-weighted average AI-system execution cost and\nAI-system completion time (GPT-OSS 120B costs about 3.1% more and is about 2.1% faster), yet GPT-OSS 120B uses\n68% less energy. Despite its lower net value, GPT-OSS 120B delivers more per\nmegajoule. By contrast, GLM 5.2 delivers more per dollar of AI cost and per\nAI hour. This reversal\nis possible because execution cost reflects allocated resources\nand elapsed time, and completion time records those elapsed minutes; measured\nenergy instead integrates GPU, CPU, and DRAM power during computation, memory\naccess, and waiting (<a href=\"#bib.bib16\">Chung et al., 2025</a>). Similar costs and durations can\ntherefore mask different joule totals.</p>\n<h3 id=\"3-3-how-do-accelerator-type-and-count-affect-return-per-joule-q3\">3.3 How do accelerator type and count affect return per joule? (Q3)</h3>\n<p>Qwen3.6-27B is used for the accelerator experiments because it fits on a single\naccelerator, allowing both one-accelerator sweeps and comparison with the\none-node, eight-H100 configuration above.</p>\n<p>Accelerator type materially\nchanges the observed return per joule.\nWe run Qwen3.6-27B with OpenHands on one NVIDIA B200, H200, or H100 and one AMD\nInstinct MI350X, using the same expert-only reference and judge. B200 delivers\nthe highest GDP-weighted NVJ and TSJ among these deployments\n(Figure <a href=\"#S3.F6\" title=\"Figure 6 ‣ 3.3 How do accelerator type and count affect return per joule? (Q3) ‣ 3 Experiments and Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">6</a>). With Qwen3.6-27B held fixed, differences\nin accelerator architecture, memory bandwidth, and power draw help explain\ndifferences in inference time and energy use, and hence the spread in NVJ and\nTSJ across devices.</p>\n<p>One H100 delivers more net value per joule, whereas eight deliver more per AI hour.\nAcross all tasks, moving Qwen3.6-27B from one to eight H100s reduces\nGDP-weighted mean AI-system completion time by 22% but uses 3.24 times the\nenergy while delivering nearly the same output dollar value. Eight H100s\nyield more net value per AI hour, while one H100 yields higher NVJ and TSJ\n(Figure <a href=\"#S3.F7\" title=\"Figure 7 ‣ 3.3 How do accelerator type and count affect return per joule? (Q3) ‣ 3 Experiments and Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">7</a>).</p>\n<p>Tensor parallelism splits inference across eight H100s, helping reduce\ncompletion time, but GPU communication and synchronization, together with\nsequential tool calls, limit the end-to-end gain (<a href=\"#bib.bib16\">Chung et al., 2025</a>).\nAll eight GPUs draw power throughout the run, so the shorter duration does\nnot offset their higher aggregate power use. With output value nearly\nunchanged, this explains why eight H100s improve return per AI hour while\none H100 delivers higher NVJ and TSJ.</p>\n<p>Figure 6: Accelerator choice changes return per joule. Qwen3.6-27B, one accelerator, expert-only reference, and GDP weights. Whiskers show 95% task-composition intervals.</p>\n<p>Figure 7: Eight H100s deliver similar value but lower return per joule than one. Qwen3.6-27B, expert-only reference, and GDP weights. Whiskers show 95% task-composition intervals. Appendix <a href=\"#A13\" title=\"Appendix M Accelerator Comparisons ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">M</a> gives the full comparison.</p>\n<h2 id=\"4-limitations-and-future-work\">4 Limitations and Future Work</h2>\n<p>Benchmark scope and evaluation.\nFollowing GDPval, output dollar value uses expert completion time and occupation\nmedian wage, not measured GDP. GDP-contributing occupations and GDP-based weights\nmotivate the framing; future work could validate the proxy with independent\nexpert valuations and test broader task samples.</p>\n<p>System coverage.\nControlled accelerator sweeps vary hardware type and count. Future comparisons\ncould vary model, harness, backend, precision, batching, and retries individually\nto isolate their effects and test portability beyond OpenHands.</p>\n<p>Measurement and deployment.\nEnergy covers metered GPU, CPU, and DRAM; execution cost uses published rates\n(Appendix <a href=\"#A7\" title=\"Appendix G Calculating AI-System Execution Cost ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">G</a>). Deployment studies could meter facility\noverhead and estimate expert review, coordination, liability, and adoption costs.</p>\n<p>Toward optimization.\nFuture work can test NVJ and TSJ as training signals and use them for routing\nand stopping. Such policies should spend computation only when expected value\nor time saving warrants the energy, subject to quality and reliability.</p>\n<h2 id=\"5-conclusion\">5 Conclusion</h2>\n<p>Joules-to-GDP connects judged professional work with the energy used to\nproduce it. The evaluated tasks and systems illustrate a framework rather than\na comprehensive survey. In deployments with limited energy, this accounting\nmakes computation allocation an explicit choice across tasks, models, and\naccelerators. Comparing total benefit with return per joule can then guide\nrouting, stopping, and serving decisions toward higher economic value and\ngreater expert time savings within that budget.</p>\n<h2 id=\"reproducibility-statement\">Reproducibility Statement</h2>\n<p>We will release our task-level evaluation dataset with expert annotations,\nblinded model judgments, completion times, execution costs, and energy\nmeasurements. The GDPval task dataset is public, and the five evaluated models\nhave publicly available weights. The release will identify the model and judge\nversions and give access details for any restricted reference materials.</p>\n<p>We will also release the study-authored harness configuration,\nenergy-accounting and aggregation code, figure and table builders, fixed\nanalysis seeds, and a source inventory with version identifiers and hashes.\nThese materials support regeneration of reported results from the same recorded\ninputs. The methods and appendix document expert references, judging, run\nselection, cost boundaries, measured energy components, economic weights, and\nvalidation checks. End-to-end reruns require compatible hardware and serving\nsoftware and may produce different outputs or energy measurements.</p>\n<h2 id=\"acknowledgements\">Acknowledgements</h2>\n<p>We thank the Scaling Intelligence Lab and Hazy Lab for constructive feedback on this paper. We are especially grateful to Snorkel AI for its feedback and for providing the materials and input that made this research possible. At Snorkel AI, we thank Ramya Ramakrishnan, Vincent Sunn Chen, Srikar Kodati, Amanda Dsouza, and Paroma Varma. We thank Yasa Baig, Rishi Bommasani, Catherine Dong, Yuzhen Mao, Shayan Talaei, and Thoughtful Lab for their feedback. We thank Modal, the Laude Institute, Google Cloud Platform, IBM Research, and Stanford HAI for providing compute resources that enabled our experiments. We also gratefully acknowledge the support of NSF under Nos. CCF2247015 (Hardware-Aware) and 2522743 (AIMing); DARPA under No. HR00112520038 (Fallingwater); NSTXL under No. N00164-23-9-G057-01 (Energy-Efficient AI Hardware); ONR under No. N000142312633 (Deep Signal Processing); Stanford HAI under No. 345077 (Evo); Google DeepMind; Google Research; Google Cloud; NXP, Xilinx, LETI-CEA, Intel, IBM (Code Generation), Microsoft, NEC, Toshiba, TSMC, ARM, Hitachi, BASF, Accenture, Ericsson, Qualcomm, Analog Devices, Salesforce, Total, Lightspeed Venture Partners, the Laude Institute, Prime Intellect, Anthropic, the HAI-GCP Cloud Credits for Research program, the Stanford Data Science Initiative (SDSI), Knight-Hennessy Scholarship, JP Morgan AI/ML Fellowship, Stanford EDGE Fellowship, GEM Fellowship, and members of the Stanford DAWN project: Meta, Google, and VMWare; and members of the Stanford SEAMS project: IBM and Felicis Ventures. We also gratefully acknowledge computing resources provided by Stanford’s Marlowe Computing Instrument. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright notation thereon. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views, policies, or endorsements, either expressed or implied, of NSF, ONR, DARPA, or the U.S. Government.</p>\n<h2 id=\"references\">References</h2>\n<ul><li><p>Acemoglu (2024)</p><p>D. Acemoglu\nThe simple macroeconomics of AI.\nTechnical report\nTechnical Report 32487, National Bureau of Economic Research.\nCited by: <a href=\"#A1.p2.1\" title=\"Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">Appendix A</a>.</p></li><li><p>Amazon Web Services (2026a)</p><p>Amazon Web Services\nAmazon EBS pricing.\nNote: <a href=\"https://aws.amazon.com/ebs/pricing/\" rel=\"nofollow ugc noopener\">https://aws.amazon.com/ebs/pricing/</a>Accessed September 19, 2026\nCited by: <a href=\"#A7.SS0.SSS0.Px1.p1.1\" title=\"Rates and resource boundary. ‣ Appendix G Calculating AI-System Execution Cost ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">Appendix G</a>.</p></li><li><p>Amazon Web Services (2026b)</p><p>Amazon Web Services\nAmazon EC2 On-Demand pricing and data transfer.\nNote: <a href=\"https://aws.amazon.com/ec2/pricing/on-demand/\" rel=\"nofollow ugc noopener\">https://aws.amazon.com/ec2/pricing/on-demand/</a>Accessed September 19, 2026\nCited by: <a href=\"#A7.SS0.SSS0.Px1.p1.1\" title=\"Rates and resource boundary. ‣ Appendix G Calculating AI-System Execution Cost ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">Appendix G</a>.</p></li><li><p>Amazon Web Services (2026c)</p><p>Amazon Web Services\nAmazon EC2 P5 instances.\nNote: <a href=\"https://aws.amazon.com/ec2/instance-types/p5/\" rel=\"nofollow ugc noopener\">https://aws.amazon.com/ec2/instance-types/p5/</a>Accessed September 19, 2026\nCited by: <a href=\"#A7.SS0.SSS0.Px1.p1.1\" title=\"Rates and resource boundary. ‣ Appendix G Calculating AI-System Execution Cost ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">Appendix G</a>.</p></li><li><p>Amazon Web Services (2026d)</p><p>Amazon Web Services\nAmazon EC2 versioned price list, us-east-1, version 20260728175247.\nNote: <a href=\"https://pricing.us-east-1.amazonaws.com/offers/v1.0/aws/AmazonEC2/20260728175247/us-east-1/index.csv\" rel=\"nofollow ugc noopener\">https://pricing.us-east-1.amazonaws.com/offers/v1.0/aws/AmazonEC2/20260728175247/us-east-1/index.csv</a>Accessed September 19, 2026\nCited by: <a href=\"#A7.SS0.SSS0.Px1.p1.1\" title=\"Rates and resource boundary. ‣ Appendix G Calculating AI-System Execution Cost ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">Appendix G</a>,\n<a href=\"#A7.SS1.p1.2\" title=\"G.1 Execution-cost sensitivity ‣ Appendix G Calculating AI-System Execution Cost ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">§G.1</a>.</p></li><li><p>Amazon Web Services (2026e)</p><p>Amazon Web Services\nAmazon VPC pricing: public IPv4 addresses.\nNote: <a href=\"https://aws.amazon.com/vpc/pricing/\" rel=\"nofollow ugc noopener\">https://aws.amazon.com/vpc/pricing/</a>Accessed September 19, 2026\nCited by: <a href=\"#A7.SS0.SSS0.Px1.p1.1\" title=\"Rates and resource boundary. ‣ Appendix G Calculating AI-System Execution Cost ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">Appendix G</a>.</p></li><li><p>Backlund and Petersson (2025)</p><p>A. Backlund and L. Petersson\nVending-Bench: a benchmark for long-term coherence of autonomous agents.\narXiv preprint arXiv:2502.15840.\nCited by: <a href=\"#A1.p1.1\" title=\"Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">Appendix A</a>.</p></li><li><p>Bonney et al. (2026)</p><p>K. Bonney, C. Breaux, E. Dinlersoz, L. Foster, J. Haltiwanger, and A. Pande\nThe microstructure of AI diffusion: evidence from firms, business functions, and worker tasks.\nTechnical report\nTechnical Report CES-26-25, U.S. Census Bureau, Center for Economic Studies.\nCited by: <a href=\"#A1.p2.1\" title=\"Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">Appendix A</a>.</p></li><li><p>Brynjolfsson et al. (2025)</p><p>E. Brynjolfsson, D. Li, and L. R. Raymond\nGenerative AI at work.\nThe Quarterly Journal of Economics 140 (2), pp. 889–942.\nCited by: <a href=\"#A1.p1.1\" title=\"Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">Appendix A</a>.</p></li><li><p>Chung et al. (2025)</p><p>J. Chung, J. J. Ma, R. Wu, J. Liu, O. J. Kweon, Y. Xia, Z. Wu, and M. Chowdhury\nThe ML.ENERGY benchmark: toward automated inference energy measurement and optimization.\narXiv preprint arXiv:2505.06371.\nCited by: <a href=\"#A1.p3.1\" title=\"Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">Appendix A</a>,\n<a href=\"#S1.p1.1\" title=\"1 Introduction ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">§1</a>,\n<a href=\"#S3.SS2.SSS1.p2.1\" title=\"3.2.1 Net value per dollar of AI cost and per AI hour ‣ 3.2 How do models trade off economic value and energy? (Q2) ‣ 3 Experiments and Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">§3.2.1</a>,\n<a href=\"#S3.SS3.p4.1\" title=\"3.3 How do accelerator type and count affect return per joule? (Q3) ‣ 3 Experiments and Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">§3.3</a>.</p></li><li><p>CloudRift (2026)</p><p>CloudRift\nGPU rental pricing and included resources.\nNote: <a href=\"https://www.cloudrift.ai/pricing\" rel=\"nofollow ugc noopener\">https://www.cloudrift.ai/pricing</a>Accessed September 19, 2026\nCited by: <a href=\"#A7.SS0.SSS0.Px1.p2.1\" title=\"Rates and resource boundary. ‣ Appendix G Calculating AI-System Execution Cost ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">Appendix G</a>.</p></li><li><p>CoreWeave (2026)</p><p>CoreWeave\nNVIDIA HGX H200 pricing.\nNote: <a href=\"https://www.coreweave.com/pricing\" rel=\"nofollow ugc noopener\">https://www.coreweave.com/pricing</a>Accessed September 2026\nCited by: <a href=\"#A7.SS1.p1.2\" title=\"G.1 Execution-cost sensitivity ‣ Appendix G Calculating AI-System Execution Cost ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">§G.1</a>.</p></li><li><p>Deloitte (2026)</p><p>Deloitte\nDeloitte’s enterprise AI infrastructure survey: a 2028 outlook.\nNote: <a href=\"https://www.deloitte.com/us/en/insights/topics/technology-management/ai-infrastructure-survey.html\" rel=\"nofollow ugc noopener\">https://www.deloitte.com/us/en/insights/topics/technology-management/ai-infrastructure-survey.html</a>Survey of 515 U.S. business and technology leaders, fielded November–December 2025\nCited by: <a href=\"#S1.p8.1\" title=\"1 Introduction ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">§1</a>.</p></li><li><p>Eloundou et al. (2023)</p><p>T. Eloundou, S. Manning, P. Mishkin, and D. Rock\nGPTs are GPTs: an early look at the labor market impact potential of large language models.\narXiv preprint arXiv:2303.10130.\nCited by: <a href=\"#A1.p2.1\" title=\"Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">Appendix A</a>.</p></li><li><p>Erol et al. (2026)</p><p>M. H. Erol, B. El, M. Suzgun, M. Yuksekgonul, and J. Zou\nCost-of-pass: an economic framework for evaluating language models.\nIn International Conference on Learning Representations,\nCited by: <a href=\"#A1.p3.1\" title=\"Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">Appendix A</a>,\n<a href=\"#S1.p1.1\" title=\"1 Introduction ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">§1</a>.</p></li><li><p>Felten et al. (2021)</p><p>E. W. Felten, M. Raj, and R. Seamans\nOccupational, industry, and geographic exposure to artificial intelligence: a novel dataset and its potential uses.\nStrategic Management Journal 42 (12).\nCited by: <a href=\"#A1.p2.1\" title=\"Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">Appendix A</a>.</p></li><li><p>Google Cloud (2026)</p><p>Google Cloud\nAccelerator-optimized machine family pricing.\nNote: <a href=\"https://cloud.google.com/products/compute/pricing/accelerator-optimized\" rel=\"nofollow ugc noopener\">https://cloud.google.com/products/compute/pricing/accelerator-optimized</a>Accessed September 20, 2026\nCited by: <a href=\"#A7.SS1.p1.2\" title=\"G.1 Execution-cost sensitivity ‣ Appendix G Calculating AI-System Execution Cost ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">§G.1</a>.</p></li><li><p>Google DeepMind (2026a)</p><p>Google DeepMind\nGemini 3.1 Flash-Lite model card.\nNote: Official model documentation\nCited by: <a href=\"#S2.SS2.p1.1\" title=\"2.2 Blinded evaluation ‣ 2 Methods ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">§2.2</a>.</p></li><li><p>Google DeepMind (2026b)</p><p>Google DeepMind\nGemma 4 model card and Apache 2.0 license.\nNote: <a href=\"https://ai.google.dev/gemma/docs/core/model_card_4\" rel=\"nofollow ugc noopener\">https://ai.google.dev/gemma/docs/core/model_card_4</a>Accessed September 2026\nCited by: <a href=\"#A3.p2.1\" title=\"Appendix C Benchmark Contamination ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">Appendix C</a>,\n<a href=\"#S3.SS2.p1.1\" title=\"3.2 How do models trade off economic value and energy? (Q2) ‣ 3 Experiments and Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">§3.2</a>.</p></li><li><p>Handa et al. (2025)</p><p>K. Handa A. Tamkin et al.\nWhich economic tasks are performed with AI? evidence from millions of Claude conversations.\narXiv preprint arXiv:2503.04761.\nCited by: <a href=\"#A1.p2.1\" title=\"Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">Appendix A</a>.</p></li><li><p>Hazy Research (2026a)</p><p>Hazy Research\nBenchmarking intelligence efficiency of LM inference.\nNote: <a href=\"https://github.com/HazyResearch/intelligence-per-watt\" rel=\"nofollow ugc noopener\">https://github.com/HazyResearch/intelligence-per-watt</a>Accessed September 2026\nCited by: <a href=\"#A5.p2.1\" title=\"Appendix E Execution and End-to-End Energy Accounting ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">Appendix E</a>.</p></li><li><p>Hazy Research (2026b)</p><p>Hazy Research\nTelemetry overview.\nNote: <a href=\"https://hazyresearch.stanford.edu/intelligence-per-watt/benchmarking/overview/\" rel=\"nofollow ugc noopener\">https://hazyresearch.stanford.edu/intelligence-per-watt/benchmarking/overview/</a>Accessed September 22, 2026\nCited by: <a href=\"#A5.p2.1\" title=\"Appendix E Execution and End-to-End Energy Accounting ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">Appendix E</a>.</p></li><li><p>Jha et al. (2025)</p><p>S. Jha et al.\nITBench: evaluating AI agents across diverse real-world IT automation tasks.\nIn Proceedings of the 42nd International Conference on Machine Learning,\nProceedings of Machine Learning Research, Vol. 267.\nCited by: <a href=\"#A1.p1.1\" title=\"Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">Appendix A</a>.</p></li><li><p>Kapoor et al. (2024)</p><p>S. Kapoor et al.\nAI agents that matter.\narXiv preprint arXiv:2407.01502.\nCited by: <a href=\"#A1.p3.1\" title=\"Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">Appendix A</a>.</p></li><li><p>Kapoor et al. (2025)</p><p>S. Kapoor, B. Stroebl, P. Kirgis, N. Nadgir, Z. S. Siegel, B. Wei, T. Xue, Z. Chen, et al.\nHolistic agent leaderboard: the missing infrastructure for AI agent evaluation.\narXiv preprint arXiv:2510.11977.\nCited by: <a href=\"#A1.p3.1\" title=\"Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">Appendix A</a>.</p></li><li><p>Korinek and McKelvey (2026)</p><p>A. Korinek and P. McKelvey\nMeasuring the AI economy.\nTechnical report\nTechnical Report Staff Working Paper 2026-20, Bank of Canada.\nCited by: <a href=\"#A1.p2.1\" title=\"Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">Appendix A</a>.</p></li><li><p>Krupp et al. (2026)</p><p>L. Krupp, D. Geißler, V. Banwari, P. Lukowicz, and J. Karolus\nPromoting sustainable web agents: benchmarking and estimating energy consumption through empirical and theoretical analysis.\nProceedings of the AAAI Conference on Artificial Intelligence 40 (45), pp. 38772–38780.\nExternal Links: <a href=\"https://dx.doi.org/10.1609/aaai.v40i45.41221\" rel=\"nofollow ugc noopener\">Document</a>,\n<a href=\"https://ojs.aaai.org/index.php/AAAI/article/view/41221\" rel=\"nofollow ugc noopener\">Link</a>\nCited by: <a href=\"#A1.p3.1\" title=\"Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">Appendix A</a>.</p></li><li><p>Luccioni et al. (2024)</p><p>A. S. Luccioni, Y. Jernite, and E. Strubell\nPower hungry processing: watts driving the cost of AI deployment?.\nIn Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency,\npp. 85–99.\nCited by: <a href=\"#A1.p3.1\" title=\"Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">Appendix A</a>.</p></li><li><p>Mazeika et al. (2025)</p><p>M. Mazeika, A. Gatti, C. Menghini, U. M. Sehwag, S. Singhal, Y. Orlovskiy, et al.\nRemote labor index: measuring AI automation of remote work.\narXiv preprint arXiv:2510.26787.\nCited by: <a href=\"#A1.p1.1\" title=\"Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">Appendix A</a>.</p></li><li><p>Modal (2026)</p><p>Modal\nResource pricing: GPU, CPU and memory.\nNote: <a href=\"https://modal.com/pricing\" rel=\"nofollow ugc noopener\">https://modal.com/pricing</a>Accessed September 19, 2026\nCited by: <a href=\"#A7.SS0.SSS0.Px1.p2.1\" title=\"Rates and resource boundary. ‣ Appendix G Calculating AI-System Execution Cost ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">Appendix G</a>.</p></li><li><p>National Renewable Energy Laboratory (2011)</p><p>National Renewable Energy Laboratory\nPlug load assessment guidelines.\nTechnical report\nTechnical Report NREL/FS-7A20-50130, National Renewable Energy Laboratory.\nExternal Links: <a href=\"https://docs.nlr.gov/docs/fy11osti/50130.pdf\" rel=\"nofollow ugc noopener\">Link</a>\nCited by: <a href=\"#footnote2\" title=\"In 3.2 How do models trade off economic value and energy? (Q2) ‣ 3 Experiments and Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">footnote 2</a>.</p></li><li><p>Niu et al. (2026)</p><p>C. Niu, W. Zhang, J. Li, Y. Zhao, T. Wang, X. Wang, and Y. Chen\nTokenPowerBench: benchmarking the power consumption of LLM inference.\nProceedings of the AAAI Conference on Artificial Intelligence 40 (38), pp. 32582–32590.\nExternal Links: <a href=\"https://dx.doi.org/10.1609/aaai.v40i38.40535\" rel=\"nofollow ugc noopener\">Document</a>,\n<a href=\"https://ojs.aaai.org/index.php/AAAI/article/view/40535\" rel=\"nofollow ugc noopener\">Link</a>\nCited by: <a href=\"#A1.SS0.SSS0.Px1.p1.1\" title=\"From token efficiency to economic return. ‣ Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">Appendix A</a>.</p></li><li><p>Noy and Zhang (2023)</p><p>S. Noy and W. Zhang\nExperimental evidence on the productivity effects of generative artificial intelligence.\nScience 381 (6654), pp. 187–192.\nCited by: <a href=\"#A1.p1.1\" title=\"Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">Appendix A</a>.</p></li><li><p>NVIDIA (2026)</p><p>NVIDIA\nNVIDIA H200 Tensor Core GPU: specifications.\nNote: Official product specificationsH200 SXM maximum configurable TDP: 700 W. Accessed 2026-09-22\nExternal Links: <a href=\"https://www.nvidia.com/en-us/data-center/h200/\" rel=\"nofollow ugc noopener\">Link</a>\nCited by: <a href=\"#S3.SS2.p2.1\" title=\"3.2 How do models trade off economic value and energy? (Q2) ‣ 3 Experiments and Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">§3.2</a>.</p></li><li><p>O*NET Resource Center (2019)</p><p>O*NET Resource Center\nO*NET-SOC 2019 taxonomy.\nNote: National Center for O*NET Development\nExternal Links: <a href=\"https://www.onetcenter.org/taxonomy/2019/\" rel=\"nofollow ugc noopener\">Link</a>\nCited by: <a href=\"#A8.p1.1\" title=\"Appendix H Economic Weights and Wage Robustness ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">Appendix H</a>.</p></li><li><p>OpenAI (2025a)</p><p>OpenAI\ngpt-oss model card and Apache 2.0 license.\nNote: <a href=\"https://huggingface.co/openai/gpt-oss-120b\" rel=\"nofollow ugc noopener\">https://huggingface.co/openai/gpt-oss-120b</a>Accessed September 2026\nCited by: <a href=\"#S3.SS2.p1.1\" title=\"3.2 How do models trade off economic value and energy? (Q2) ‣ 3 Experiments and Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">§3.2</a>.</p></li><li><p>OpenAI (2025b)</p><p>OpenAI\ngpt-oss-120b &amp; gpt-oss-20b model card.\nNote: Official model documentation\nCited by: <a href=\"#A3.p2.1\" title=\"Appendix C Benchmark Contamination ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">Appendix C</a>.</p></li><li><p>OpenAI (2025c)</p><p>OpenAI\nMeasuring the performance of our models on real-world tasks.\nNote: <a href=\"https://openai.com/index/gdpval/\" rel=\"nofollow ugc noopener\">https://openai.com/index/gdpval/</a>GDPval public announcement\nCited by: <a href=\"#A3.p2.1\" title=\"Appendix C Benchmark Contamination ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">Appendix C</a>.</p></li><li><p>Patwardhan et al. (2025)</p><p>T. Patwardhan, R. Dias, E. Proehl, G. Kim, M. Wang, O. Watkins, S. Posada Fishman, M. Aljubeh, P. Thacker, L. Fauconnet, N. S. Kim, P. Chao, S. Miserendino, G. Chabot, D. Li, M. Sharman, A. Barr, A. Glaese, and J. Tworek\nGDPval: evaluating AI model performance on real-world economically valuable tasks.\narXiv preprint arXiv:2510.04374.\nCited by: <a href=\"#A1.p1.1\" title=\"Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">Appendix A</a>,\n<a href=\"#A8.p1.1\" title=\"Appendix H Economic Weights and Wage Robustness ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">Appendix H</a>,\n<a href=\"#S1.p1.1\" title=\"1 Introduction ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">§1</a>,\n<a href=\"#S2.SS1.p1.1\" title=\"2.1 Dataset and expert references ‣ 2 Methods ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">§2.1</a>,\n<a href=\"#S2.SS1.p4.1\" title=\"2.1 Dataset and expert references ‣ 2 Methods ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">§2.1</a>,\n<a href=\"#S2.SS2.p1.1\" title=\"2.2 Blinded evaluation ‣ 2 Methods ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">§2.2</a>.</p></li><li><p>Qwen Team (2026)</p><p>Qwen Team\nQwen3.6-27B model card and Apache 2.0 license.\nNote: <a href=\"https://huggingface.co/Qwen/Qwen3.6-27B\" rel=\"nofollow ugc noopener\">https://huggingface.co/Qwen/Qwen3.6-27B</a>Accessed September 2026\nCited by: <a href=\"#A3.p2.1\" title=\"Appendix C Benchmark Contamination ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">Appendix C</a>,\n<a href=\"#S3.SS2.p1.1\" title=\"3.2 How do models trade off economic value and energy? (Q2) ‣ 3 Experiments and Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">§3.2</a>.</p></li><li><p>Saad-Falcon et al. (2025)</p><p>J. Saad-Falcon, A. Narayan, H. O. Akengin, J. W. Griffin, H. Shandilya, A. Gamarra Lafuente, M. Goel, R. Joseph, S. Natarajan, E. K. Guha, S. Zhu, B. Athiwaratkun, J. Hennessy, A. Mirhoseini, and C. Ré\nIntelligence per watt: measuring intelligence efficiency of local AI.\narXiv preprint arXiv:2511.07885.\nCited by: <a href=\"#A1.p3.1\" title=\"Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">Appendix A</a>,\n<a href=\"#S2.SS3.p1.1\" title=\"2.3 End-to-end energy accounting ‣ 2 Methods ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">§2.3</a>.</p></li><li><p>Samsi et al. (2023)</p><p>S. Samsi, D. Zhao, J. McDonald, B. Li, A. Michaleas, M. Jones, W. Bergeron, J. Kepner, D. Tiwari, and V. Gadepally\nFrom words to watts: benchmarking the energy costs of large language model inference.\narXiv preprint arXiv:2310.03003.\nExternal Links: <a href=\"https://arxiv.org/abs/2310.03003\" rel=\"nofollow ugc noopener\">Link</a>\nCited by: <a href=\"#A1.SS0.SSS0.Px1.p1.1\" title=\"From token efficiency to economic return. ‣ Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">Appendix A</a>.</p></li><li><p>Shao et al. (2025)</p><p>Y. Shao, H. Zope, Y. Jiang, J. Pei, D. Nguyen, E. Brynjolfsson, and D. Yang\nFuture of work with AI agents: auditing automation and augmentation potential across the U.S. workforce.\narXiv preprint arXiv:2506.06576.\nCited by: <a href=\"#A1.p2.1\" title=\"Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">Appendix A</a>.</p></li><li><p>Starace et al. (2025)</p><p>G. Starace et al.\nPaperBench: evaluating AI’s ability to replicate AI research.\nIn Proceedings of the 42nd International Conference on Machine Learning,\nProceedings of Machine Learning Research, Vol. 267.\nCited by: <a href=\"#A1.p1.1\" title=\"Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">Appendix A</a>.</p></li><li><p>Tamkin and McCrory (2025)</p><p>A. Tamkin and P. McCrory\nEstimating AI productivity gains from Claude conversations.\nNote: Anthropic Economic Index\nCited by: <a href=\"#A1.p2.1\" title=\"Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">Appendix A</a>.</p></li><li><p>Tripathy et al. (2026)</p><p>A. Tripathy, C. P. Harshit, and K. Vaidhyanathan\nSWEnergy: an empirical study on energy efficiency in agentic issue resolution frameworks with SLMs.\nIn Proceedings of the 2026 International Workshop on Agentic Engineering,\npp. 104–111.\nExternal Links: <a href=\"https://dx.doi.org/10.1145/3786167.3788406\" rel=\"nofollow ugc noopener\">Document</a>,\n<a href=\"https://doi.org/10.1145/3786167.3788406\" rel=\"nofollow ugc noopener\">Link</a>\nCited by: <a href=\"#A1.p3.1\" title=\"Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">Appendix A</a>,\n<a href=\"#S1.p1.1\" title=\"1 Introduction ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">§1</a>.</p></li><li><p>Tschand et al. (2024)</p><p>A. Tschand, A. T. R. Rajan, S. Idgunji, A. Ghosh, J. Holleman, C. Kiraly, et al.\nMLPerf Power: benchmarking the energy efficiency of machine learning systems from microwatts to megawatts for sustainable AI.\narXiv preprint arXiv:2410.12032.\nCited by: <a href=\"#A1.SS0.SSS0.Px1.p1.1\" title=\"From token efficiency to economic return. ‣ Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">Appendix A</a>,\n<a href=\"#A1.p3.1\" title=\"Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">Appendix A</a>.</p></li><li><p>U.S. Bureau of Labor Statistics (2025)</p><p>U.S. Bureau of Labor Statistics\nMay 2024 national occupational employment and wage estimates.\nNote: Occupational Employment and Wage StatisticsMay 2024 estimates, released April 2025\nExternal Links: <a href=\"https://www.bls.gov/oes/2024/may/oes_nat.htm\" rel=\"nofollow ugc noopener\">Link</a>\nCited by: <a href=\"#A8.p1.1\" title=\"Appendix H Economic Weights and Wage Robustness ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">Appendix H</a>.</p></li><li><p>Vellaisamy et al. (2026)</p><p>P. Vellaisamy, V. Lam, S. Blanton, and J. P. Shen\nCharacterization of request and token energy costs for LLM inference workloads on GPU platforms.\narXiv preprint arXiv:2608.28044.\nExternal Links: <a href=\"https://arxiv.org/abs/2608.28044\" rel=\"nofollow ugc noopener\">Link</a>\nCited by: <a href=\"#A1.SS0.SSS0.Px1.p1.1\" title=\"From token efficiency to economic return. ‣ Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">Appendix A</a>.</p></li><li><p>Vidgen et al. (2026)</p><p>B. Vidgen, A. Mann, A. Fennelly, J. Wright Stanly, L. Rothman, M. Burstein, J. Benchek, D. Ostrofsky, A. Ravichandran, D. Sur, N. Venugopal, A. Hsia, I. Robinson, C. Huang, O. Varones, D. Khan, M. Haines, A. Bridges, J. Boyle, K. Twist, Z. Richards, C. Mahapatra, B. Foody, and O. Nitski\nAPEX–Agents.\nNote: arXiv preprint arXiv:2601.14242\nCited by: <a href=\"#A1.p1.1\" title=\"Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">Appendix A</a>,\n<a href=\"#A11.SS2.p5.3\" title=\"K.2 Repeated runs: Pass@𝑘 and four-run win–tie thresholds ‣ Appendix K Expert-Reference and Repeated-Run Sensitivity ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">§K.2</a>.</p></li><li><p>Wan et al. (2026)</p><p>A. Wan, S. Hatgis-Kessell, T. Aguirre, P. Liang, and R. Bommasani\nEconomic evaluations of language models.\narXiv preprint arXiv:2607.19375.\nCited by: <a href=\"#A1.p2.1\" title=\"Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">Appendix A</a>,\n<a href=\"#S1.p1.1\" title=\"1 Introduction ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">§1</a>.</p></li><li><p>Wei et al. (2025)</p><p>T. Wei, W. Wen, R. Qiao, X. Sun, and J. Ma\nRocketEval: efficient automated LLM evaluation via grading checklist.\nIn The Thirteenth International Conference on Learning Representations,\nExternal Links: <a href=\"https://openreview.net/forum?id=zJjzNj6QUe\" rel=\"nofollow ugc noopener\">Link</a>\nCited by: <a href=\"#A4.SS1.SSS0.Px2.p1.1\" title=\"Expert review. ‣ D.1 Judge consistency ‣ Appendix D Judge and Evidence Pipeline ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">§D.1</a>.</p></li><li><p>Xie et al. (2024)</p><p>T. Xie et al.\nOSWorld: benchmarking multimodal agents for open-ended tasks in real computer environments.\narXiv preprint arXiv:2404.07972.\nCited by: <a href=\"#A1.p1.1\" title=\"Appendix A Related Work ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">Appendix A</a>.</p></li><li><p>Z.ai (2026)</p><p>Z.ai\nGLM-5.2 model card and MIT license.\nNote: <a href=\"https://huggingface.co/zai-org/GLM-5.2\" rel=\"nofollow ugc noopener\">https://huggingface.co/zai-org/GLM-5.2</a>Accessed September 2026\nCited by: <a href=\"#A3.p2.1\" title=\"Appendix C Benchmark Contamination ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">Appendix C</a>,\n<a href=\"#S3.SS2.p1.1\" title=\"3.2 How do models trade off economic value and energy? (Q2) ‣ 3 Experiments and Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">§3.2</a>.</p></li></ul>\n<h2 id=\"appendix-a-related-work\">Appendix A Related Work</h2>\n<p>Professional work and economic value. <em>GDPval</em>\n(<a href=\"#bib.bib2\">Patwardhan et al., 2025</a>) evaluates deliverables;\n<em>APEX-Agents</em> and the <em>Remote Labor Index</em>\n(<a href=\"#bib.bib4\">Vidgen et al., 2026</a>; <a href=\"#bib.bib8\">Mazeika et al., 2025</a>) compare AI systems with experts\non longer tasks. Related benchmarks extend this approach to research, IT, web,\ndesktop, organizational settings, and economic simulations\n(<a href=\"#bib.bib25\">Starace and others, 2025</a>; <a href=\"#bib.bib26\">Jha and others, 2025</a>; <a href=\"#bib.bib27\">Xie and others, 2024</a>; <a href=\"#bib.bib5\">Backlund and Petersson, 2025</a>). Controlled and field studies complement these benchmarks\nby measuring changes in workers’ completion time, output quality, and productivity\n(<a href=\"#bib.bib6\">Noy and Zhang, 2023</a>; <a href=\"#bib.bib7\">Brynjolfsson et al., 2025</a>).</p>\n<p>Occupations and the wider economy. <em>GPTs Are GPTs</em>\n(<a href=\"#bib.bib28\">Eloundou et al., 2023</a>) estimates which occupational tasks language models\ncould affect. The <em>Anthropic Economic Index</em>\n(<a href=\"#bib.bib29\">Handa et al., 2025</a>; <a href=\"#bib.bib9\">Tamkin and McCrory, 2025</a>) uses real Claude conversations\nto study where AI is being used at work and estimate possible time saving,\nwhile <em>Economic Evaluations of Language Models</em> (<a href=\"#bib.bib30\">Wan et al., 2026</a>) tests\ntasks across the U.S. occupational economy and estimates how much work time current\nmodels could save. Related studies connect AI capabilities, use, and adoption\nto workers, occupations, firms, industries, and the broader economy\n(<a href=\"#bib.bib31\">Felten et al., 2021</a>; <a href=\"#bib.bib10\">Shao et al., 2025</a>; <a href=\"#bib.bib11\">Bonney et al., 2026</a>; <a href=\"#bib.bib13\">Acemoglu, 2024</a>; <a href=\"#bib.bib12\">Korinek and McKelvey, 2026</a>).</p>\n<p>Cost and energy efficiency. *AI Agents\nThat Matter*, <em>Cost-of-Pass</em>, and the <em>Holistic Agent Leaderboard</em> compare\nperformance with financial or usage-based costs\n(<a href=\"#bib.bib32\">Kapoor and others, 2024</a>; <a href=\"#bib.bib14\">Erol et al., 2026</a>; <a href=\"#bib.bib15\">Kapoor et al., 2025</a>). Energy studies such as\n<em>Power Hungry Processing</em>, <em>ML.ENERGY</em>, and <em>MLPerf Power</em> directly measure energy\nuse across models, workloads, and accelerators\n(<a href=\"#bib.bib17\">Luccioni et al., 2024</a>; <a href=\"#bib.bib16\">Chung et al., 2025</a>; <a href=\"#bib.bib18\">Tschand et al., 2024</a>; <a href=\"#bib.bib19\">Saad-Falcon et al., 2025</a>). More recent\nwork, including <em>SWEnergy</em> and <em>Sustainable Web Agents</em>, extends energy measurement\nto AI-system workflows (<a href=\"#bib.bib33\">Tripathy et al., 2026</a>; <a href=\"#bib.bib34\">Krupp et al., 2026</a>).\nNVJ and TSJ extend this resource accounting to net dollar value and time saving\nrather than benchmark scores or win–tie rates alone.</p>\n<h5 id=\"from-token-efficiency-to-economic-return\">From token efficiency to economic return.</h5>\n<p>Token-level studies vary batch size, context length, and GPU count, including\nphase-specific prefill and decoding measurements\n(<a href=\"#bib.bib49\">Samsi et al., 2023</a>; <a href=\"#bib.bib50\">Niu et al., 2026</a>). Longer generations can lower\njoules per token by amortizing prefill while increasing total request energy\n(<a href=\"#bib.bib51\">Vellaisamy et al., 2026</a>); long contexts enlarge attention and\nkey–value-cache traffic. For a common workload, window, and energy boundary,\ntokens per joule multiplied by net value or time saving per output token\ngives NVJ or TSJ, after unit conversion. Token efficiency therefore complements\noutput quality, task value, and time saving; tool execution and waiting also\nconsume energy without producing model tokens. Comparing these measures requires\nmatched boundaries because GPU counters and MLPerf Power’s external whole-system\nmetering cover different components (<a href=\"#bib.bib18\">Tschand et al., 2024</a>).</p>\n<h2 id=\"appendix-b-expert-references-and-annotations\">Appendix B Expert References and Annotations</h2>\n<p>The tasks span heterogeneous formats, including documents,\nspreadsheets, presentations, code, images, audio, video, and archive-based\ndeliverables. The frozen task records also preserve occupation, sector, and\nO*NET-SOC identifiers; source-file and deliverable inventories; and the\ngrading rubric. The expert-only reference was produced by an expert without\nAI; its recorded time is decomposed into prompt review,\nsource-file use, task execution, and verification or quality review.\nThe expert-with-AI reference was produced by an expert using AI. Its records\nadd the AI provider, model version, interface, source files\nused, external tools, full transcripts, and turn-level\nexpert-active, AI-generation-wait, and total times.\nThe supplied assisted-reference records identify Claude Sonnet 4.6 and\nClaude Opus 4.7 used through Claude Web Chat.\nThese are recorded expert completion times. The assisted condition includes\nthe expert’s setup, monitoring, review, and correction, as well as AI waits.\nThe expert-only and expert-with-AI references were annotated by different people.</p>\n<p>The pairwise comparison with an expert deliverable assesses relative judged\nquality; it does not establish absolute task quality or realized economic\nsavings. We use the GDPval expert deliverables as meaningful reference points\nwithout assuming they are error-free.</p>\n<p>The rubric helps ground the\ncomparison in the task requirements even when an expert deliverable has flaws.\nExpert-only reference, expert-with-AI reference, and\nAI-system deliverables are normalized to the same evidence representation.\nCandidate position is balanced exactly\nacross the archived task-condition comparisons.</p>\n<p>The expert-with-AI reference is a supplied assisted completion condition;\nit is not a claim about every possible form of assisted professional work.\nIts completion time can exceed expert-only time on individual tasks.\nThe expert-only reference is the cleaner\ncommon reference for autonomous systems because neither its\ndeliverables nor its output dollar values depend on a particular assisting model,\ninterface, or expert–AI workflow. We therefore retain it as primary; the\nexpert-with-AI reference is a secondary, workflow-specific\noperational comparison. Appendix <a href=\"#A11.SS1\" title=\"K.1 Separating reference judgments from time valuation ‣ Appendix K Expert-Reference and Repeated-Run Sensitivity ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">K.1</a> separates\nchanges in judged win–tie outcomes from changes in time valuation.</p>\n<h2 id=\"appendix-c-benchmark-contamination\">Appendix C Benchmark Contamination</h2>\n<p>During evaluation, the models received each task and its required source files,\nbut not the grading rubric or expert deliverable. We checked AI deliverables\nfor exact matches to expert files and unexplained identical document or\nworkbook components and found none.</p>\n<p>GPT-OSS 120B and GPT-OSS 20B were released\nbefore GDPval became public (<a href=\"#bib.bib23\">OpenAI, 2025b</a>; <a href=\"#bib.bib3\">OpenAI, 2025c</a>).\nThe public model cards for GLM 5.2, Qwen3.6-27B, and Gemma 4 31B do not\nmention GDPval\n(<a href=\"#bib.bib44\">Z.ai, 2026</a>; <a href=\"#bib.bib45\">Qwen Team, 2026</a>; <a href=\"#bib.bib46\">Google DeepMind, 2026b</a>).\nThese checks show no direct copying in the evaluated outputs; they do not\nestablish every model’s training-data contents.</p>\n<h2 id=\"appendix-d-judge-and-evidence-pipeline\">Appendix D Judge and Evidence Pipeline</h2>\n<p>The judging pipeline is model-agnostic and can be configured with\nalternative language-model judges. Main economic results use the\nfrozen judge models/gemini-3.1-flash-lite at temperature 0. Its input\ncontains the task specification, grading rubric, source evidence, and both\nidentity-masked candidate deliverables. The evidence pipeline uses full sanitized\nsemantic text and rendered visual PDFs with a 48-page visual ceiling. It\nperforms safe archive extraction, static code inspection, and media parsing.\nAll verdict records used in the analysis parsed without error.</p>\n<h3 id=\"d-1-judge-consistency\">D.1 Judge consistency</h3>\n<p>Table <a href=\"#A4.T2\" title=\"Table 2 ‣ D.1 Judge consistency ‣ Appendix D Judge and Evidence Pipeline ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">2</a> compares Gemini 3.1 Flash-Lite (the main\njudge), Claude Haiku 4.5, and GPT-5 mini on the same GLM 5.2 comparisons,\nincluding all four attempts per task.</p>\n<p>Table 2: The three judges agree on 72–77% of exact outcomes. An AI win or tie receives task credit. Exact agreement\nretains win/tie/loss; binary win–tie agreement collapses wins and ties into one category. All comparisons use GLM 5.2 outputs from four attempts per task.</p>\n<div class=\"table-wrap\"><table><thead><tr><th>Pair</th><th>Exact agreement</th><th>Binary win–tie agreement</th></tr></thead><tbody><tr><td>Gemini–Haiku</td><td>77.2%</td><td>79.2%</td></tr><tr><td>Gemini–GPT-5 mini</td><td>74.0%</td><td>76.4%</td></tr><tr><td>Haiku–GPT-5 mini</td><td>72.3%</td><td>72.8%</td></tr></tbody></table></div>\n<h5 id=\"repeated-output-agreement\">Repeated-output agreement.</h5>\n<p>For GLM 5.2, the first two independently generated outputs for each of 220\ntasks produced 440 Gemini judgments. The win–tie versus loss decision matched\nacross the two outputs for 171 tasks (77.7%).</p>\n<h5 id=\"expert-review\">Expert review.</h5>\n<p>In a blinded expert review of 32 randomly selected GLM 5.2 task comparisons,\nexperts agreed with Gemini 3.1 Flash-Lite on 23 of 32 win–tie versus loss\ndecisions (71.9%). A separate automated review yielded the same agreement\nrate. Human-to-human agreement on related pairwise response judgments has\nbeen reported at 64.7% (<a href=\"#bib.bib1\">Wei et al., 2025</a>).</p>\n<h2 id=\"appendix-e-execution-and-end-to-end-energy-accounting\">Appendix E Execution and End-to-End Energy Accounting</h2>\n<p>The AI-system runs were executed in concurrent batches. IPW telemetry\nsampled the metered components at a configured 0.2-second\ninterval. The end-to-end measurement window began immediately before the\nbatch’s runs launched and ended when the final run completed. For\ncomponents exposing cumulative energy counters, end-to-end energy used the\nboundary-aligned difference between the end and start readings. Otherwise,\nsampled power was integrated over the same wall-clock interval. Measurements\nfrom all accelerators assigned to a run were summed over the execution.</p>\n<p>For every evaluated deployment, IPW measured GPU, CPU, and DRAM energy.\nReported joules sum these component-energy readings over the execution\nwindow (<a href=\"#bib.bib48\">Hazy Research, 2026a</a>; <a href=\"#bib.bib52\">Hazy Research, 2026b</a>).\nTable <a href=\"#A5.T3\" title=\"Table 3 ‣ Appendix E Execution and End-to-End Energy Accounting ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">3</a> gives the serving settings.</p>\n<p>Table 3: All evaluated deployments use vLLM, 20 concurrent requests, a 131,072-token context limit, and a 32,768-token output limit. Tensor parallelism is eight for eight-accelerator deployments and one for single-accelerator deployments.</p>\n<div class=\"table-wrap\"><table><thead><tr><th>Model and accelerator</th><th>Inference</th><th>Tensor parallelism</th><th>Concurrent requests</th><th>Context limit (tokens)</th><th>Output limit (tokens)</th></tr></thead><tbody><tr><td>GLM 5.2, 8×\\timesH200</td><td>vLLM</td><td>TP8</td><td>20</td><td>131,072</td><td>32,768</td></tr><tr><td>Qwen3.6-27B, 8×\\timesH100</td><td>vLLM</td><td>TP8</td><td>20</td><td>131,072</td><td>32,768</td></tr><tr><td>Gemma 4 31B, 8×\\timesH100</td><td>vLLM</td><td>TP8</td><td>20</td><td>131,072</td><td>32,768</td></tr><tr><td>GPT-OSS 120B, 8×\\timesH100</td><td>vLLM</td><td>TP8</td><td>20</td><td>131,072</td><td>32,768</td></tr><tr><td>GPT-OSS 20B, 8×\\timesH100</td><td>vLLM</td><td>TP8</td><td>20</td><td>131,072</td><td>32,768</td></tr><tr><td>Qwen3.6-27B, 1×\\timesB200</td><td>vLLM</td><td>TP1</td><td>20</td><td>131,072</td><td>32,768</td></tr><tr><td>Qwen3.6-27B, 1×\\timesH200</td><td>vLLM</td><td>TP1</td><td>20</td><td>131,072</td><td>32,768</td></tr><tr><td>Qwen3.6-27B, 1×\\timesH100</td><td>vLLM</td><td>TP1</td><td>20</td><td>131,072</td><td>32,768</td></tr><tr><td>Qwen3.6-27B, 1×\\timesMI350X</td><td>vLLM</td><td>TP1</td><td>20</td><td>131,072</td><td>32,768</td></tr></tbody></table></div>\n<p>We chose OpenHands as the primary harness because its compact tool surface, a\nterminal and file editor, was sufficient for producing and validating the\nheterogeneous deliverables in GDPval. The run ledger records turns, tool\ncalls, AI-system completion time, outcome, and energy, allowing the same evaluation protocol to\nbe applied to other harnesses and inference backends. The codebase also includes\nportability paths for MiniMax and Apple M4; these\nconditions are not pooled with the main results unless the complete evaluation\nprotocol is matched.</p>\n<p>The active-request allocation in Equation <a href=\"#S2.E2\" title=\"In 2.3 End-to-end energy accounting ‣ 2 Methods ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">2</a> reconciles\nto the retained batch totals. This conservation identity applies to a full\nbatch; it does not imply that concurrent requests draw equal instantaneous\npower.</p>\n<h2 id=\"appendix-f-energy-attribution-sensitivity\">Appendix F Energy-Attribution Sensitivity</h2>\n<p>A concurrent batch has one measured energy total, while individual requests\noverlap in time. Our primary active-request rule\n(Equation <a href=\"#S2.E2\" title=\"In 2.3 End-to-end energy accounting ‣ 2 Methods ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">2</a>) assigns each moment’s energy across the\nrequests then running. An equal-batch-share alternative assigns the same\nfraction of the batch total to every request. Both rules conserve measured\nbatch energy, but they can assign different energy to individual tasks\nbecause requests finish at different times.</p>\n<p>We apply both rules to the evaluated GLM 5.2 task pool underlying\nFigure <a href=\"#S3.F3\" title=\"Figure 3 ‣ 3 Experiments and Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">3</a>, holding AI outputs, judge decisions,\nexpert times, wages, AI-system execution costs, and GDP weights fixed.\nTable <a href=\"#A6.T4\" title=\"Table 4 ‣ Appendix F Energy-Attribution Sensitivity ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">4</a> gives the category comparisons.\nProtective Service remains first in win–tie rate per MJ, and Computer &amp;\nMathematical remains first in NVJ. The TSJ leader changes. Architecture &amp; Engineering leads\nunder the primary active-request rule, while Computer &amp; Mathematical leads\nunder equal batch shares. Eight NVJ ranks and nine TSJ ranks change across\nthe 13 categories. Both rules conserve each measured batch total; neither establishes\nthe energy drawn by an individual concurrent request.</p>\n<p>Figure 8: Request-completion share does not track batch energy proportionally. P20, P50, P80, and P90 mark completion of 20%, 50%, 80%, and 90% of requests. Each point pools batch totals after estimating energy consumed at that milestone. Intermediate points use each batch’s average power and recorded completion times; the 100% endpoint is measured batch energy. The dotted line shows proportional completion for reference.</p>\n<p>Table 4: Equal batch shares change the TSJ leader but leave the NVJ leader unchanged. Each cell shows active-request allocation / equal batch share. Active-request allocation is the paper’s primary method. AI outputs, judgments, expert times, wages, costs, and GDP weights are fixed; only energy assignment changes. Values are point estimates.</p>\n<div class=\"table-wrap\"><table><thead><tr><th>Category</th><th>Energy (MJ)</th><th>NVJ (USD/MJ)</th><th>TSJ (hours/MJ)</th></tr></thead><tbody><tr><td>Computer &amp; Mathematical</td><td>0.355 / 0.337</td><td>523.8 / 551.6</td><td>7.29 / 7.67</td></tr><tr><td>Architecture &amp; Engineering</td><td>0.495 / 0.597</td><td>413.9 / 343.2</td><td>7.89 / 6.54</td></tr><tr><td>Management</td><td>0.360 / 0.349</td><td>374.6 / 387.0</td><td>6.75 / 6.97</td></tr><tr><td>Protective Service</td><td>0.225 / 0.229</td><td>278.1 / 272.9</td><td>4.96 / 4.86</td></tr><tr><td>Production</td><td>0.550 / 0.786</td><td>221.1 / 154.6</td><td>6.12 / 4.28</td></tr><tr><td>Community &amp; Social Service</td><td>0.282 / 0.363</td><td>188.5 / 146.3</td><td>6.16 / 4.78</td></tr><tr><td>Legal</td><td>0.370 / 0.334</td><td>149.7 / 166.1</td><td>1.50 / 1.66</td></tr><tr><td>Business &amp; Financial Operations</td><td>0.505 / 0.481</td><td>145.6 / 152.9</td><td>3.32 / 3.49</td></tr><tr><td>Personal Care &amp; Service</td><td>0.401 / 0.503</td><td>134.3 / 107.2</td><td>7.15 / 5.70</td></tr><tr><td>Arts, Design, Entertainment &amp; Media</td><td>0.450 / 0.427</td><td>103.8 / 109.5</td><td>2.71 / 2.86</td></tr><tr><td>Healthcare Practitioners &amp; Technical</td><td>0.445 / 0.504</td><td>97.8 / 86.3</td><td>1.67 / 1.47</td></tr><tr><td>Sales &amp; Related</td><td>0.356 / 0.362</td><td>65.4 / 64.4</td><td>2.01 / 1.98</td></tr><tr><td>Office &amp; Administrative Support</td><td>0.604 / 0.567</td><td>40.1 / 42.6</td><td>1.84 / 1.96</td></tr></tbody></table></div>\n<p>Figure <a href=\"#A6.F8\" title=\"Figure 8 ‣ Appendix F Energy-Attribution Sensitivity ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">8</a> describes when energy is consumed during\nthe measured execution batches. For each batch, we find when a given fraction\nof its requests has finished and multiply elapsed batch time by its measured\naverage power. P20, P50, P80, and P90 correspond to an estimated 37.9%,\n52.4%, 71.1%, and 83.9% of pooled batch energy. These intermediate values\nuse a constant-power approximation; only the full-batch energy total is\ndirectly measured. The figure does not evaluate a stopping policy, since\nenergy already spent on unfinished requests is included at each milestone\nand the outcomes of stopping them are not observed.</p>\n<h2 id=\"appendix-g-calculating-ai-system-execution-cost\">Appendix G Calculating AI-System Execution Cost</h2>\n<p>We calculate AI-system execution cost for the evaluated runs from recorded\nGPU, CPU, and RAM use at published rates. Those rates include hardware and\nelectricity, so energy is not charged again in dollars.</p>\n<h5 id=\"rates-and-resource-boundary\">Rates and resource boundary.</h5>\n<p>The eight-H100 comparisons use the July 2026 AWS Linux On-Demand\np5.48xlarge rate of $55.04 per node-hour. GLM 5.2 uses the\np5en.48xlarge rate of $63.296 per node-hour for eight H200s.\nThese are whole-node rates including host CPU and RAM, not rates per GPU.\nThe calculation also includes one boot volume, one public IPv4 address, and\none final artifact export where applicable\n(<a href=\"#bib.bib35\">Amazon Web Services, 2026d</a>; <a href=\"#bib.bib36\">Amazon Web Services, 2026c</a>; <a href=\"#bib.bib37\">Amazon Web Services, 2026a</a>; <a href=\"#bib.bib38\">Amazon Web Services, 2026e</a>; <a href=\"#bib.bib39\">Amazon Web Services, 2026b</a>).</p>\n<p>For the single-NVIDIA-accelerator comparison, we use Modal’s published\nper-second H100, H200, and B200 prices together with a common 32-core,\n256-GiB host reservation (<a href=\"#bib.bib40\">Modal, 2026</a>). The MI350X comparison\nuses CloudRift’s published on-demand GPU-hour rate\n(<a href=\"#bib.bib43\">CloudRift, 2026</a>). These provider prices do not guarantee\nidentical host specifications or performance across providers.</p>\n<h5 id=\"time-allocation-and-setup\">Time allocation and setup.</h5>\n<p>Let Ai​(t)A_{i}(t) indicate when request ii is active, and let N⁡(t)N(t)\ncount active requests. We allocate shared active node time using\nτi=∫Ai​(t)/N⁡(t)​𝑑t\\tau_{i}=\\int A_{i}(t)/N(t)\\,dt. Shared startup, gaps, and finalization\ntime T0T_{0} inside a task-execution window contributes\nT0​τi/∑jτjT_{0}\\tau_{i}/\\sum_{j}\\tau_{j}, so the recorded window is charged once.\nThus the total node time allocated to request ii is\nTialloc=τi+T0​τi/∑jτjT_{i}^{\\mathrm{alloc}}=\\tau_{i}+T_{0}\\tau_{i}/\\sum_{j}\\tau_{j}.\nThe comparison represents a prepared local service; initialization outside\nthe task-execution window is excluded from the reported execution cost.</p>\n<h5 id=\"scope\">Scope.</h5>\n<p>Published prices can change. The calculation excludes unobserved\nexternal API charges, durable storage, expert review, development,\ncoordination, and adoption costs. These belong in a broader deployment\nassessment alongside the measured energy and the value of delivered work.</p>\n<h3 id=\"g-1-execution-cost-sensitivity\">G.1 Execution-cost sensitivity</h3>\n<p>For each evaluated GLM 5.2 task, we calculate the AI-system execution cost by\nmultiplying its allocated node seconds by the AWS node-hour rate divided by\n3,600, then adding its assigned boot-volume and IPv4 charges. To examine a\nhypothetical alternative provider price, we replace only that hourly node\nrate. For task (i), the resulting cost at rate (r) is</p>\n<div class=\"table-wrap\"><table><thead><tr><th></th><th></th><th></th><th></th></tr></thead><tbody><tr><td></td><td>Ci​(r)=Ci​(AWS)+Tialloc3,600​(r−63.296),C_{i}(r)=C_{i}(\\mathrm{AWS})+\\frac{T_{i}^{\\mathrm{alloc}}}{3{,}600}(r-63.296),</td><td></td><td>(6)</td></tr></tbody></table></div>\n<p>where TiallocT_{i}^{\\mathrm{alloc}} includes active and allocated shared node seconds. We use published eight-H200\nnode tariffs of $50.44/hour for CoreWeave and $84.8069/hour for Google Cloud\nalongside the $63.296/hour AWS rate\n(<a href=\"#bib.bib42\">CoreWeave, 2026</a>; <a href=\"#bib.bib41\">Google Cloud, 2026</a>; <a href=\"#bib.bib35\">Amazon Web Services, 2026d</a>).\nThe same recorded outcomes, time, energy, and non-node charges are held fixed;\nthe alternative-provider columns are hypothetical price calculations, not\nruns on those providers. Table <a href=\"#A7.T5\" title=\"Table 5 ‣ G.1 Execution-cost sensitivity ‣ Appendix G Calculating AI-System Execution Cost ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">5</a> shows\nthe resulting category costs and NVJ. The category NVJ order is unchanged\nat these three tariffs, while TSJ is unchanged because execution cost does\nnot enter its numerator or denominator.</p>\n<p>Table 5: Published node tariffs change calculated costs without changing category NVJ order. For the same recorded GLM 5.2 tasks, only the node-hour price is changed; task outcomes, time, energy, and non-node charges remain fixed. Values are hypothetical costs per category-normalized task and the resulting NVJ, not results of executions on other providers. The AWS column reproduces the evaluated cost boundary.</p>\n<div class=\"table-wrap\"><table><thead><tr><th></th><th>AI execution cost (USD)</th><th></th><th></th><th>NVJ (USD/MJ)</th><th></th></tr></thead><tbody><tr><td>Category</td><td>CoreWeave</td><td>AWS</td><td>GCP</td><td>CoreWeave</td><td>GCP</td></tr><tr><td>Architecture &amp; Eng.</td><td>1.677</td><td>2.104</td><td>2.819</td><td>414.8</td><td>412.5</td></tr><tr><td>Arts &amp; Media</td><td>1.582</td><td>1.986</td><td>2.660</td><td>104.7</td><td>102.3</td></tr><tr><td>Business &amp; Financial</td><td>1.731</td><td>2.172</td><td>2.911</td><td>146.4</td><td>144.1</td></tr><tr><td>Community &amp; Social</td><td>0.983</td><td>1.233</td><td>1.652</td><td>189.4</td><td>187.0</td></tr><tr><td>Computer &amp; Math</td><td>1.245</td><td>1.562</td><td>2.092</td><td>524.6</td><td>522.3</td></tr><tr><td>Healthcare</td><td>1.837</td><td>2.306</td><td>3.089</td><td>98.9</td><td>96.1</td></tr><tr><td>Legal</td><td>1.257</td><td>1.577</td><td>2.113</td><td>150.6</td><td>148.3</td></tr><tr><td>Management</td><td>1.256</td><td>1.576</td><td>2.112</td><td>375.5</td><td>373.1</td></tr><tr><td>Office &amp; Admin.</td><td>2.173</td><td>2.727</td><td>3.653</td><td>41.0</td><td>38.5</td></tr><tr><td>Personal Care</td><td>1.506</td><td>1.890</td><td>2.532</td><td>135.3</td><td>132.7</td></tr><tr><td>Production</td><td>1.859</td><td>2.333</td><td>3.125</td><td>222.0</td><td>219.7</td></tr><tr><td>Protective Service</td><td>0.778</td><td>0.976</td><td>1.308</td><td>279.0</td><td>276.6</td></tr><tr><td>Sales</td><td>1.265</td><td>1.587</td><td>2.127</td><td>66.3</td><td>63.9</td></tr></tbody></table></div>\n<h2 id=\"appendix-h-economic-weights-and-wage-robustness\">Appendix H Economic Weights and Wage Robustness</h2>\n<p>GDPval chose sectors for their GDP contribution and sampled predominantly\nknowledge-work occupations with large aggregate compensation (<a href=\"#bib.bib2\">Patwardhan et al., 2025</a>).\nIts public gold set has five tasks per occupation, so equal task weights would\nreflect the sampling design rather than sector economic scale. We give each of\nthe nine sectors its normalized GDP share, then divide that share among its\nevaluated occupations in proportion to national employment. Sector GDP thus\nsets the scale across sectors, while employment distinguishes occupations\nwithin each sector. We do not use wage-bill weights because wages already enter\noutput dollar value as recorded expert completion time multiplied by the\noccupation’s matched OEWS median hourly wage (<a href=\"#bib.bib21\">U.S. Bureau of Labor Statistics, 2025</a>; <a href=\"#bib.bib2\">Patwardhan et al., 2025</a>).\nUsing wages again in the weights would give high-wage occupations additional\ninfluence; wage-bill weights would instead emphasize compensation paid to those\noccupations. The GDP-based weights preserve sector GDP shares without assuming\nequal GDP-to-wage-bill ratios across sectors. The 44 occupations are grouped into 13\noccupation categories using the O*NET-SOC taxonomy (<a href=\"#bib.bib22\">O*NET Resource Center, 2019</a>);\nO*NET-SOC supplies the classification and does not construct the weights.</p>\n<p>Let γs\\gamma_{s} be sector ss’s share of U.S. GDP, normalized across the\nnine selected sectors, and let\nLo=EmploymentoL_{o}=\\mathrm{Employment}_{o}. The occupation\nweight used in both aggregation levels is</p>\n<div class=\"table-wrap\"><table><thead><tr><th></th><th></th><th></th><th></th></tr></thead><tbody><tr><td></td><td>woGDP=γs⁡(o)Lo∑j:s⁡(j)=s⁡(o)Lj,∑o∈𝒪woGDP=1.w_{o}^{\\mathrm{GDP}}=\\gamma_{s(o)}\\frac{L_{o}}{\\sum_{j:s(j)=s(o)}L_{j}},\\qquad\\sum_{o\\in\\mathcal{O}}w_{o}^{\\mathrm{GDP}}=1.</td><td></td><td>(7)</td></tr></tbody></table></div>\n<p>For completeness, define the per-occupation task average as\nX~o=Xo/no\\widetilde{X}<em>{o}=X</em>{o}/n_{o}, where non_{o} is its evaluated task count and\nX∈{G,C,S,E}X\\in{G,C,S,E}, with SS denoting time saving as in\nSection <a href=\"#S2.SS5\" title=\"2.5 Two levels of GDP-based aggregation ‣ 2 Methods ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">2.5</a>. For category cc, its category weight and within-category\noccupation weights are</p>\n<div class=\"table-wrap\"><table><thead><tr><th></th><th></th><th></th><th></th></tr></thead><tbody><tr><td></td><td>Wc=∑o∈cwoGDP,wo</td><td>c=woGDPWc,∑o∈cwo</td><td>c=1.W_{c}=\\sum_{o\\in c}w_{o}^{\\mathrm{GDP}},\\qquad w_{o\\mid c}=\\frac{w_{o}^{\\mathrm{GDP}}}{W_{c}},\\qquad\\sum_{o\\in c}w_{o\\mid c}=1.</td></tr></tbody></table></div>\n<p>The category outcome is\nX~c=∑o∈cwo|c​X~o\\widetilde{X}<em>{c}=\\sum</em>{o\\in c}w_{o\\mid c}\\widetilde{X}_{o}, while the\nGDP-weighted benchmark outcome is\nX~GDP=∑cWc​X~c\\widetilde{X}^{\\mathrm{GDP}}=\\sum_{c}W_{c}\\widetilde{X}_{c}, with\n∑cWc=1\\sum_{c}W_{c}=1. The same within-category and benchmark weights are applied to\noutput dollar value, AI-system execution cost, time saving and energy before\nforming NVJ, TSJ or net value per dollar of AI cost; the method never\naverages already-formed occupation or category ratios.</p>\n<p>Table <a href=\"#A8.T6\" title=\"Table 6 ‣ Appendix H Economic Weights and Wage Robustness ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">6</a> lists the resulting category weights, and\nFigure <a href=\"#A8.F9\" title=\"Figure 9 ‣ Appendix H Economic Weights and Wage Robustness ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">9</a> visualizes their construction and use.</p>\n<p>Figure 9: Wages determine output dollar value without also increasing benchmark weights. Sector GDP shares and within-sector employment shares determine occupation weights. Category comparisons renormalize those weights to one within each category, while the GDP-weighted benchmark metric retains category weights. Expert completion time and occupation-specific median wages determine output dollar value.</p>\n<p>Table 6: Sales and Management receive the largest GDP-based category weights. Employment shares allocate each sector’s GDP share among its evaluated occupations; the category weights sum to one and are retained in the GDP-weighted benchmark aggregate.</p>\n<p>Category\nOccupations\nWeight\nEmployment (M)</p>\n<p>Management\n7\n19.68%\n6.77</p>\n<p>Business &amp; Financial Ops.\n6\n12.44%\n3.95</p>\n<p>Computer &amp; Mathematical\n1\n3.22%\n1.65</p>\n<p>Architecture &amp; Engineering\n2\n3.16%\n0.64</p>\n<p>Community &amp; Social Service\n1\n3.82%\n0.38</p>\n<p>Legal\n1\n1.45%\n0.75</p>\n<p>Arts, Design, Entertainment &amp; Media\n5\n7.13%\n0.38</p>\n<p>Healthcare Practitioners &amp; Technical\n3\n6.10%\n3.92</p>\n<p>Protective Service\n2\n1.59%\n0.19</p>\n<p>Personal Care &amp; Service\n2\n3.91%\n0.35</p>\n<p>Sales &amp; Related\n8\n20.24%\n4.00</p>\n<p>Office &amp; Administrative Support\n5\n13.87%\n5.99</p>\n<p>Production\n1\n3.39%\n0.69</p>\n<p>Table <a href=\"#A8.T7\" title=\"Table 7 ‣ Appendix H Economic Weights and Wage Robustness ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">7</a> compares mean and median hourly wages in task\nvaluation while holding GDP-based occupation weights, energy, AI-system execution cost, and\ntime saving fixed. The headline analysis uses median hourly wages.\nU.S. GDP shares, employment, wages, work practices, and estimates of time saving\nchange over time; results should therefore be recomputed as work changes.</p>\n<p>Table 7: Mean rather than median wages raise NVJ while leaving TSJ unchanged. AI-system execution cost is held fixed for GDP-weighted GLM 5.2 with one attempt per task. NVJ is USD/MJ; TSJ is hours/MJ.</p>\n<p>Wage basis</p>\n<p>Output dollar\nvalue (USD)</p>\n<p>AI-system\nexecution cost (USD)\nEnergy (MJ)\nNVJ\nTSJ</p>\n<p>OEWS mean\n85.57\n1.92\n0.430\n194.58\n3.823</p>\n<p>OEWS median\n74.19\n1.92\n0.430\n168.11\n3.823</p>\n<h2 id=\"appendix-i-occupation-category-results\">Appendix I Occupation-Category Results</h2>\n<p>Table <a href=\"#A9.T8\" title=\"Table 8 ‣ Appendix I Occupation-Category Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">8</a> gives win–tie rates,\nTable <a href=\"#A9.T9\" title=\"Table 9 ‣ Appendix I Occupation-Category Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">9</a> gives net dollar value and time saving\nper task, and Table <a href=\"#A9.T10\" title=\"Table 10 ‣ Appendix I Occupation-Category Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">10</a> gives returns per MJ.\nThese tables underlie Figure <a href=\"#S3.F3\" title=\"Figure 3 ‣ 3 Experiments and Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">3</a>. Brackets give 95% task-composition\nintervals. They describe variation across the evaluated tasks and do not\nestablish a population ordering of categories.</p>\n<p>Table 8: Protective Service leads win–tie rate per MJ, while Personal Care leads win–tie rate. GLM 5.2, expert-only reference, one attempt per task; occupation outcomes are category-normalized with GDP-based weights. Brackets show 95% task-composition intervals.</p>\n<div class=\"table-wrap\"><table><thead><tr><th>Category</th><th>Win–tie (%)</th><th>Win–tie pp/MJ</th></tr></thead><tbody><tr><td>Architecture &amp; Eng.</td><td>69.00 [38.01, 91.00]</td><td>139.43 [75.84, 229.50]</td></tr><tr><td>Arts &amp; Media</td><td>46.09 [28.27, 66.09]</td><td>102.44 [74.15, 136.56]</td></tr><tr><td>Business &amp; Financial</td><td>56.81 [37.93, 74.97]</td><td>112.42 [72.81, 168.62]</td></tr><tr><td>Community &amp; Social</td><td>60.00 [20.00, 100.00]</td><td>212.94 [60.19, 390.26]</td></tr><tr><td>Computer &amp; Math</td><td>80.00 [40.00, 100.00]</td><td>225.08 [122.22, 312.77]</td></tr><tr><td>Healthcare</td><td>28.23 [8.02, 63.12]</td><td>63.48 [21.77, 107.89]</td></tr><tr><td>Legal</td><td>60.00 [20.00, 100.00]</td><td>162.14 [49.98, 321.76]</td></tr><tr><td>Management</td><td>78.67 [62.10, 92.17]</td><td>218.37 [162.96, 289.27]</td></tr><tr><td>Office &amp; Admin.</td><td>38.14 [16.10, 62.36]</td><td>63.18 [23.59, 135.67]</td></tr><tr><td>Personal Care</td><td>95.79 [87.37, 100.00]</td><td>238.63 [150.53, 515.22]</td></tr><tr><td>Production</td><td>40.00 [0.00, 80.00]</td><td>72.78 [0.00, 150.11]</td></tr><tr><td>Protective Service</td><td>60.00 [20.80, 99.20]</td><td>267.01 [84.52, 479.38]</td></tr><tr><td>Sales</td><td>55.31 [34.84, 74.31]</td><td>155.28 [96.17, 222.56]</td></tr></tbody></table></div>\n<p>Table 9: Architecture &amp; Engineering leads net dollar value and time saving per task. GLM 5.2, expert-only reference, one attempt per task; occupation outcomes are category-normalized with GDP-based weights. Brackets show 95% task-composition intervals.</p>\n<div class=\"table-wrap\"><table><thead><tr><th>Category</th><th>Net dollar value (USD/task)</th><th>Time saving (hours/task)</th></tr></thead><tbody><tr><td>Architecture &amp; Eng.</td><td>204.84 [108.93, 288.94]</td><td>3.905 [2.055, 5.569]</td></tr><tr><td>Arts &amp; Media</td><td>46.71 [20.18, 79.15]</td><td>1.219 [0.554, 2.028]</td></tr><tr><td>Business &amp; Financial</td><td>73.57 [44.46, 106.85]</td><td>1.677 [0.975, 2.508]</td></tr><tr><td>Community &amp; Social</td><td>53.12 [12.78, 93.45]</td><td>1.735 [0.423, 3.047]</td></tr><tr><td>Computer &amp; Math</td><td>186.16 [68.11, 337.13]</td><td>2.589 [0.924, 4.844]</td></tr><tr><td>Healthcare</td><td>43.50 [14.21, 86.93]</td><td>0.742 [0.230, 1.541]</td></tr><tr><td>Legal</td><td>55.40 [7.06, 117.50]</td><td>0.555 [0.046, 1.338]</td></tr><tr><td>Management</td><td>134.95 [73.68, 229.83]</td><td>2.431 [1.239, 4.337]</td></tr><tr><td>Office &amp; Admin.</td><td>24.18 [5.74, 48.62]</td><td>1.111 [0.315, 2.200]</td></tr><tr><td>Personal Care</td><td>53.92 [34.42, 73.33]</td><td>2.870 [1.753, 3.975]</td></tr><tr><td>Production</td><td>121.51 [-2.22, 250.73]</td><td>3.361 [0.000, 6.915]</td></tr><tr><td>Protective Service</td><td>62.49 [7.21, 150.74]</td><td>1.114 [0.128, 2.792]</td></tr><tr><td>Sales</td><td>23.29 [15.22, 31.58]</td><td>0.717 [0.455, 0.987]</td></tr></tbody></table></div>\n<p>Table 10: Computer &amp; Mathematical leads NVJ; Architecture &amp; Engineering leads TSJ. GLM 5.2, expert-only reference, one attempt per task; occupation outcomes are category-normalized with GDP-based weights. Brackets show 95% task-composition intervals.</p>\n<div class=\"table-wrap\"><table><thead><tr><th>Category</th><th>NVJ (USD/MJ)</th><th>TSJ (hours/MJ)</th></tr></thead><tbody><tr><td>Architecture &amp; Eng.</td><td>413.90 [196.34, 744.61]</td><td>7.890 [3.660, 14.397]</td></tr><tr><td>Arts &amp; Media</td><td>103.82 [55.09, 167.63]</td><td>2.710 [1.459, 4.508]</td></tr><tr><td>Business &amp; Financial</td><td>145.57 [83.22, 239.58]</td><td>3.318 [1.820, 5.627]</td></tr><tr><td>Community &amp; Social</td><td>188.50 [43.10, 382.42]</td><td>6.158 [1.426, 12.522]</td></tr><tr><td>Computer &amp; Math</td><td>523.76 [235.93, 770.14]</td><td>7.285 [3.182, 11.080]</td></tr><tr><td>Healthcare</td><td>97.84 [38.77, 196.72]</td><td>1.669 [0.643, 3.049]</td></tr><tr><td>Legal</td><td>149.70 [21.88, 370.36]</td><td>1.500 [0.144, 4.257]</td></tr><tr><td>Management</td><td>374.57 [193.82, 672.56]</td><td>6.748 [3.253, 12.692]</td></tr><tr><td>Office &amp; Admin.</td><td>40.05 [8.22, 105.52]</td><td>1.840 [0.470, 4.745]</td></tr><tr><td>Personal Care</td><td>134.32 [87.59, 261.20]</td><td>7.151 [4.532, 14.139]</td></tr><tr><td>Production</td><td>221.10 [-4.24, 479.79]</td><td>6.116 [0.000, 13.337]</td></tr><tr><td>Protective Service</td><td>278.09 [30.55, 669.23]</td><td>4.956 [0.533, 12.268]</td></tr><tr><td>Sales</td><td>65.40 [42.16, 93.84]</td><td>2.014 [1.283, 2.898]</td></tr></tbody></table></div>\n<p>Table <a href=\"#A9.T11\" title=\"Table 11 ‣ Appendix I Occupation-Category Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">11</a> reports output dollar value and time saving\nper win or tie. These components explain why a category can yield fewer wins\nand ties per joule while delivering greater dollar value or time saving per\njoule.</p>\n<p>Table 11: Win–tie values and expert times explain why the category rankings differ. Values condition on outputs judged a win or tie for GLM 5.2 against the expert-only reference. Occupation outcomes are category-normalized with GDP-based weights.</p>\n<div class=\"table-wrap\"><table><thead><tr><th>Category</th><th>Output dollar value (USD)</th><th>Expert time (h)</th><th>AI time (h)</th><th>Time saving (h)</th></tr></thead><tbody><tr><td>Architecture &amp; Eng.</td><td>299.90</td><td>6.11</td><td>0.45</td><td>5.66</td></tr><tr><td>Arts &amp; Media</td><td>105.66</td><td>3.02</td><td>0.37</td><td>2.65</td></tr><tr><td>Business &amp; Financial</td><td>133.31</td><td>3.34</td><td>0.39</td><td>2.95</td></tr><tr><td>Community &amp; Social</td><td>90.58</td><td>3.22</td><td>0.32</td><td>2.89</td></tr><tr><td>Computer &amp; Math</td><td>234.65</td><td>3.67</td><td>0.43</td><td>3.24</td></tr><tr><td>Healthcare</td><td>162.29</td><td>3.04</td><td>0.41</td><td>2.63</td></tr><tr><td>Legal</td><td>94.96</td><td>1.31</td><td>0.38</td><td>0.93</td></tr><tr><td>Management</td><td>173.53</td><td>3.39</td><td>0.30</td><td>3.09</td></tr><tr><td>Office &amp; Admin.</td><td>70.55</td><td>3.30</td><td>0.38</td><td>2.91</td></tr><tr><td>Personal Care</td><td>58.26</td><td>3.40</td><td>0.40</td><td>3.00</td></tr><tr><td>Production</td><td>309.61</td><td>9.04</td><td>0.64</td><td>8.40</td></tr><tr><td>Protective Service</td><td>105.78</td><td>2.13</td><td>0.27</td><td>1.86</td></tr><tr><td>Sales</td><td>44.98</td><td>1.63</td><td>0.34</td><td>1.30</td></tr></tbody></table></div>\n<h2 id=\"appendix-j-why-occupation-rankings-differ\">Appendix J Why Occupation Rankings Differ</h2>\n<p>Across the evaluated tasks, let\nai|c=wo⁡(i)|c/no⁡(i)a_{i\\mid c}=w_{o(i)\\mid c}/n_{o(i)} be task ii’s normalized\nGDP-based weight within category cc, and let Yi=1Y_{i}=1 when the AI output\nwins or ties and zero otherwise. Let EiE_{i} and CiC_{i} denote measured\nenergy and AI-system execution cost. The category win–tie rate and its\nenergy-normalized version are</p>\n<div class=\"table-wrap\"><table><thead><tr><th></th><th></th><th></th><th></th><th></th></tr></thead><tbody><tr><td></td><td>pc\\displaystyle p_{c}</td><td>=∑i∈cai</td><td>c​Yi,E¯c=∑i∈cai</td><td>c​Ei,\\displaystyle=\\sum_{i\\in c}a_{i\\mid c}Y_{i},\\qquad\\bar{E}<em>{c}=\\sum</em>{i\\in c}a_{i\\mid c}E_{i},</td></tr><tr><td></td><td></td><td></td><td></td><td></td></tr><tr><td>---</td><td>---</td><td>---</td><td>---</td><td>---</td></tr><tr><td></td><td>WTc\\displaystyle\\mathrm{WT}_{c}</td><td>=100pc(%),WT/MJc=100​pcE¯c.\\displaystyle=100p_{c}\\ (\\%),\\qquad\\mathrm{WT/MJ}<em>{c}=\\frac{100p</em>{c}}{\\bar{E}_{c}}.</td><td></td><td>(10)</td></tr></tbody></table></div>\n<p>The latter unit is percentage points per MJ. All attempts contribute energy\nand cost, including losses.</p>\n<p>For pc&gt;0p_{c}&gt;0, let G¯S,c\\bar{G}_{S,c} be output dollar value per win or tie and\nD¯S,c\\bar{D}_{S,c} be time saving per win or tie. With\nC¯c=∑iai|c​Ci\\bar{C}<em>{c}=\\sum</em>{i}a_{i\\mid c}C_{i}, the two economic returns decompose as</p>\n<div class=\"table-wrap\"><table><thead><tr><th></th><th></th><th></th><th></th></tr></thead><tbody><tr><td></td><td>NVJc=pcE¯c​G¯S,c−C¯cE¯c,TSJc=pcE¯c​D¯S,c.\\mathrm{NVJ}<em>{c}=\\frac{p</em>{c}}{\\bar{E}<em>{c}}\\bar{G}</em>{S,c}-\\frac{\\bar{C}<em>{c}}{\\bar{E}</em>{c}},\\qquad\\mathrm{TSJ}<em>{c}=\\frac{p</em>{c}}{\\bar{E}<em>{c}}\\bar{D}</em>{S,c}.</td><td></td><td>(11)</td></tr></tbody></table></div>\n<p>Thus win–tie frequency per joule, dollar value per win or tie, and time\nsaving per win or tie play different roles. The category values\nare in Tables <a href=\"#A9.T8\" title=\"Table 8 ‣ Appendix I Occupation-Category Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">8</a>,\n<a href=\"#A9.T10\" title=\"Table 10 ‣ Appendix I Occupation-Category Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">10</a>, and <a href=\"#A9.T11\" title=\"Table 11 ‣ Appendix I Occupation-Category Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">11</a>.\nTable <a href=\"#A10.T12\" title=\"Table 12 ‣ Appendix J Why Occupation Rankings Differ ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">12</a> shows the resulting rank changes.\nProtective Service ranks first on win–tie rate per MJ but fourth on NVJ and\nseventh on TSJ. Computer &amp; Mathematical ranks third on win–tie rate per MJ\nbut first on NVJ because its dollar value per win or tie is higher. Production ranks\neleventh on win–tie rate per MJ and sixth on TSJ, despite the long expert\ncompletion times of its tested tasks. Architecture &amp; Engineering leads TSJ.\nEach metric therefore answers a\ndifferent question about the same attempts.</p>\n<p>Table 12: Different categories lead win–tie rate per MJ, NVJ, and TSJ. Point-estimate ranks for GLM 5.2 against the expert-only reference, with one attempt per task and category-normalized GDP-based weights; rank 1 is highest.</p>\n<div class=\"table-wrap\"><table><thead><tr><th>Category</th><th>Win–tie pp/MJ rank</th><th>NVJ rank</th><th>TSJ rank</th></tr></thead><tbody><tr><td>Architecture &amp; Eng.</td><td>8</td><td>2</td><td>1</td></tr><tr><td>Arts &amp; Media</td><td>10</td><td>10</td><td>9</td></tr><tr><td>Business &amp; Financial</td><td>9</td><td>8</td><td>8</td></tr><tr><td>Community &amp; Social</td><td>5</td><td>6</td><td>5</td></tr><tr><td>Computer &amp; Math</td><td>3</td><td>1</td><td>2</td></tr><tr><td>Healthcare</td><td>12</td><td>11</td><td>12</td></tr><tr><td>Legal</td><td>6</td><td>7</td><td>13</td></tr><tr><td>Management</td><td>4</td><td>3</td><td>4</td></tr><tr><td>Office &amp; Admin.</td><td>13</td><td>13</td><td>11</td></tr><tr><td>Personal Care</td><td>2</td><td>9</td><td>3</td></tr><tr><td>Production</td><td>11</td><td>5</td><td>6</td></tr><tr><td>Protective Service</td><td>1</td><td>4</td><td>7</td></tr><tr><td>Sales</td><td>7</td><td>12</td><td>10</td></tr></tbody></table></div>\n<h3 id=\"j-1-full-occupation-category-results\">J.1 Full Occupation-Category Results</h3>\n<p>Table <a href=\"#A10.T13\" title=\"Table 13 ‣ J.1 Full Occupation-Category Results ‣ Appendix J Why Occupation Rankings Differ ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">13</a> gives the\ncategory-normalized credited value, AI execution cost, time saving,\nand energy that form each reported ratio. Table <a href=\"#A10.T14\" title=\"Table 14 ‣ J.1 Full Occupation-Category Results ‣ Appendix J Why Occupation Rankings Differ ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">14</a>\ngives net value per dollar of AI-system execution cost and per AI hour for\nthe same categories. Table <a href=\"#A10.T15\" title=\"Table 15 ‣ J.1 Full Occupation-Category Results ‣ Appendix J Why Occupation Rankings Differ ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">15</a> gives the\ncorresponding point estimates for every evaluated occupation.\nGDP-weighting the underlying unrounded components before forming the ratios reproduces the\nGLM 5.2 row of Table <a href=\"#S3.T1\" title=\"Table 1 ‣ 3.2 How do models trade off economic value and energy? (Q2) ‣ 3 Experiments and Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">1</a> (37.69 per AI dollar\nand 191.3 USD per AI hour).</p>\n<p>Table 13: A high win–tie rate alone does not determine credited value or time saving. Category-normalized GDP-based task averages, expert-only reference. Credited value and time saving are zero on losses; AI-system execution cost and energy include every attempt. Applying (G−C)/E(G-C)/E and S/ES/E to the underlying unrounded components reproduces the NVJ and TSJ point estimates in Table <a href=\"#A9.T10\" title=\"Table 10 ‣ Appendix I Occupation-Category Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">10</a>.</p>\n<div class=\"table-wrap\"><table><thead><tr><th>Category</th><th>Win–tie (%)</th><th>Credited value GG (USD)</th><th>AI cost CC (USD)</th><th>Time saving SS (hours)</th><th>Energy EE (MJ)</th><th>AI time LL (hours)</th></tr></thead><tbody><tr><td>Architecture &amp; Eng.</td><td>69.0</td><td>206.94</td><td>2.104</td><td>3.905</td><td>0.495</td><td>0.467</td></tr><tr><td>Arts &amp; Media</td><td>46.1</td><td>48.70</td><td>1.986</td><td>1.219</td><td>0.450</td><td>0.384</td></tr><tr><td>Business &amp; Financial</td><td>56.8</td><td>75.74</td><td>2.172</td><td>1.677</td><td>0.505</td><td>0.443</td></tr><tr><td>Community &amp; Social</td><td>60.0</td><td>54.35</td><td>1.233</td><td>1.735</td><td>0.282</td><td>0.321</td></tr><tr><td>Computer &amp; Math</td><td>80.0</td><td>187.72</td><td>1.562</td><td>2.589</td><td>0.355</td><td>0.420</td></tr><tr><td>Healthcare</td><td>28.2</td><td>45.81</td><td>2.306</td><td>0.742</td><td>0.445</td><td>0.299</td></tr><tr><td>Legal</td><td>60.0</td><td>56.97</td><td>1.577</td><td>0.555</td><td>0.370</td><td>0.368</td></tr><tr><td>Management</td><td>78.7</td><td>136.53</td><td>1.576</td><td>2.431</td><td>0.360</td><td>0.319</td></tr><tr><td>Office &amp; Admin.</td><td>38.1</td><td>26.91</td><td>2.727</td><td>1.111</td><td>0.604</td><td>0.449</td></tr><tr><td>Personal Care</td><td>95.8</td><td>55.81</td><td>1.890</td><td>2.870</td><td>0.401</td><td>0.393</td></tr><tr><td>Production</td><td>40.0</td><td>123.84</td><td>2.333</td><td>3.361</td><td>0.550</td><td>0.610</td></tr><tr><td>Protective Service</td><td>60.0</td><td>63.47</td><td>0.976</td><td>1.114</td><td>0.225</td><td>0.277</td></tr><tr><td>Sales</td><td>55.3</td><td>24.88</td><td>1.587</td><td>0.717</td><td>0.356</td><td>0.326</td></tr></tbody></table></div>\n<p>Table 14: Computer &amp; Mathematical leads net value per AI dollar and per AI hour across occupation categories. GLM 5.2 against the expert-only reference, one attempt per task. Category-normalized GDP-based averages are combined before forming (G−C)/C(G-C)/C and (G−C)/L(G-C)/L, where GG is credited output dollar value, CC is AI-system execution cost, and LL is AI-system completion time in hours. All attempts contribute cost and time. Point estimates use the underlying unrounded components summarized in Table <a href=\"#A10.T13\" title=\"Table 13 ‣ J.1 Full Occupation-Category Results ‣ Appendix J Why Occupation Rankings Differ ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">13</a>; categories are ordered by net value per AI dollar.</p>\n<div class=\"table-wrap\"><table><thead><tr><th>Rank</th><th>Category</th><th>Net value/AI dollar (USD/USD)</th><th>Net value/AI hour (USD/hour)</th></tr></thead><tbody><tr><td>1</td><td>Computer &amp; Math</td><td>119.2</td><td>442.9</td></tr><tr><td>2</td><td>Architecture &amp; Eng.</td><td>97.4</td><td>438.4</td></tr><tr><td>3</td><td>Management</td><td>85.6</td><td>423.6</td></tr><tr><td>4</td><td>Protective Service</td><td>64.0</td><td>225.8</td></tr><tr><td>5</td><td>Production</td><td>52.1</td><td>199.3</td></tr><tr><td>6</td><td>Community &amp; Social</td><td>43.1</td><td>165.7</td></tr><tr><td>7</td><td>Legal</td><td>35.1</td><td>150.5</td></tr><tr><td>8</td><td>Business &amp; Financial</td><td>33.9</td><td>166.1</td></tr><tr><td>9</td><td>Personal Care</td><td>28.5</td><td>137.1</td></tr><tr><td>10</td><td>Arts &amp; Media</td><td>23.5</td><td>121.8</td></tr><tr><td>11</td><td>Healthcare</td><td>18.9</td><td>145.3</td></tr><tr><td>12</td><td>Sales</td><td>14.7</td><td>71.5</td></tr><tr><td>13</td><td>Office &amp; Admin.</td><td>8.9</td><td>53.9</td></tr></tbody></table></div>\n<p>Table 15: Occupation-level economic returns vary widely across the evaluated work. Expert-only reference and one evaluated attempt per task. Win–tie is the fraction of an occupation’s evaluated tasks judged a win or tie. Each occupation’s task components are averaged before ratios are formed. These are point estimates; all attempts contribute cost and energy.</p>\n<div class=\"table-wrap\"><table><thead><tr><th>Occupation</th><th>Tasks</th><th>Win–tie (%)</th><th>NVJ (USD/MJ)</th><th>TSJ (h/MJ)</th></tr></thead><tbody><tr><td>Accountants and Auditors</td><td>5</td><td>40.0</td><td>31.5</td><td>0.53</td></tr><tr><td>Administrative Services Managers</td><td>5</td><td>100.0</td><td>369.8</td><td>6.60</td></tr><tr><td>Audio and Video Technicians</td><td>5</td><td>40.0</td><td>22.7</td><td>0.53</td></tr><tr><td>Buyers and Purchasing Agents</td><td>5</td><td>40.0</td><td>249.3</td><td>6.69</td></tr><tr><td>Child, Family, and School Social Workers</td><td>5</td><td>60.0</td><td>188.5</td><td>6.16</td></tr><tr><td>Compliance Officers</td><td>5</td><td>80.0</td><td>333.9</td><td>8.45</td></tr><tr><td>Computer and Information Systems Managers</td><td>5</td><td>60.0</td><td>515.3</td><td>5.82</td></tr><tr><td>Concierges</td><td>5</td><td>80.0</td><td>214.2</td><td>11.18</td></tr><tr><td>Counter and Rental Clerks</td><td>5</td><td>60.0</td><td>56.4</td><td>2.71</td></tr><tr><td>Customer Service Representatives</td><td>5</td><td>40.0</td><td>45.0</td><td>2.09</td></tr><tr><td>Editors</td><td>5</td><td>80.0</td><td>407.1</td><td>10.49</td></tr><tr><td>Film and Video Editors</td><td>5</td><td>60.0</td><td>116.7</td><td>2.94</td></tr><tr><td>Financial Managers</td><td>5</td><td>40.0</td><td>40.9</td><td>0.33</td></tr><tr><td>Financial and Investment Analysts</td><td>5</td><td>40.0</td><td>231.6</td><td>4.35</td></tr><tr><td>First-Line Supervisors of Non-Retail Sales Workers</td><td>5</td><td>20.0</td><td>82.8</td><td>1.83</td></tr><tr><td>First-Line Supervisors of Office and Administrative Support Workers</td><td>5</td><td>20.0</td><td>6.1</td><td>0.18</td></tr><tr><td>First-Line Supervisors of Police and Detectives</td><td>5</td><td>60.0</td><td>292.0</td><td>5.08</td></tr><tr><td>First-Line Supervisors of Production and Operating Workers</td><td>5</td><td>40.0</td><td>221.1</td><td>6.12</td></tr><tr><td>First-Line Supervisors of Retail Sales Workers</td><td>5</td><td>60.0</td><td>65.2</td><td>2.54</td></tr><tr><td>General and Operations Managers</td><td>5</td><td>80.0</td><td>1270.8</td><td>24.74</td></tr><tr><td>Industrial Engineers</td><td>5</td><td>60.0</td><td>141.8</td><td>2.44</td></tr><tr><td>Lawyers</td><td>5</td><td>60.0</td><td>149.7</td><td>1.50</td></tr><tr><td>Mechanical Engineers</td><td>5</td><td>80.0</td><td>913.4</td><td>17.90</td></tr><tr><td>Medical Secretaries and Administrative Assistants</td><td>5</td><td>60.0</td><td>90.5</td><td>3.70</td></tr><tr><td>Medical and Health Services Managers</td><td>5</td><td>100.0</td><td>241.0</td><td>3.78</td></tr><tr><td>News Analysts, Reporters, and Journalists</td><td>5</td><td>60.0</td><td>231.4</td><td>7.59</td></tr><tr><td>Nurse Practitioners</td><td>5</td><td>80.0</td><td>653.5</td><td>9.79</td></tr><tr><td>Order Clerks</td><td>5</td><td>20.0</td><td>16.0</td><td>0.84</td></tr><tr><td>Personal Financial Advisors</td><td>5</td><td>60.0</td><td>190.4</td><td>3.29</td></tr><tr><td>Pharmacists</td><td>5</td><td>60.0</td><td>162.8</td><td>2.42</td></tr><tr><td>Private Detectives and Investigators</td><td>5</td><td>60.0</td><td>93.6</td><td>3.36</td></tr><tr><td>Producers and Directors</td><td>5</td><td>20.0</td><td>39.6</td><td>0.91</td></tr><tr><td>Project Management Specialists</td><td>5</td><td>60.0</td><td>197.1</td><td>3.52</td></tr><tr><td>Property, Real Estate, and Community Association Managers</td><td>5</td><td>80.0</td><td>119.1</td><td>2.96</td></tr><tr><td>Real Estate Brokers</td><td>5</td><td>40.0</td><td>26.9</td><td>0.58</td></tr><tr><td>Real Estate Sales Agents</td><td>5</td><td>40.0</td><td>28.1</td><td>0.73</td></tr><tr><td>Recreation Workers</td><td>5</td><td>100.0</td><td>125.4</td><td>6.70</td></tr><tr><td>Registered Nurses</td><td>5</td><td>20.0</td><td>53.4</td><td>1.05</td></tr><tr><td>Sales Managers</td><td>5</td><td>80.0</td><td>331.5</td><td>4.36</td></tr><tr><td>Sales Representatives, Wholesale and Manufacturing, Except Technical and Scientific Products</td><td>5</td><td>60.0</td><td>72.4</td><td>1.81</td></tr><tr><td>Sales Representatives, Wholesale and Manufacturing, Technical and Scientific Products</td><td>5</td><td>40.0</td><td>99.3</td><td>1.85</td></tr><tr><td>Securities, Commodities, and Financial Services Sales Agents</td><td>5</td><td>100.0</td><td>154.3</td><td>3.56</td></tr><tr><td>Shipping, Receiving, and Inventory Clerks</td><td>5</td><td>40.0</td><td>61.0</td><td>2.96</td></tr><tr><td>Software Developers</td><td>5</td><td>80.0</td><td>523.8</td><td>7.29</td></tr></tbody></table></div>\n<h3 id=\"j-2-task-composition-sensitivity\">J.2 Task-composition sensitivity</h3>\n<p>The category intervals in\nTables <a href=\"#A9.T8\" title=\"Table 8 ‣ Appendix I Occupation-Category Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">8</a> and <a href=\"#A9.T10\" title=\"Table 10 ‣ Appendix I Occupation-Category Results ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">10</a>\ncome from 10,000 task-composition bootstrap draws. Each draw resamples evaluated\ntasks within occupations, then recomputes the weighted dollar-value,\nsaved-time, cost, and energy components before forming each ratio.\nRecorded judgments and resource measurements stay fixed.</p>\n<p>We also test category-rank stability by independently removing one uniformly\nchosen task from each occupation category in 10,000 portfolios. For each\nportfolio, we re-average the retained tasks within their occupations and\nrecompute win–tie rate per MJ, NVJ, and TSJ using the same GDP occupation\nweights. This changes task composition while keeping the recorded judgments,\nAI-system execution costs, and energy measurements fixed.\nTable <a href=\"#A10.T16\" title=\"Table 16 ‣ J.2 Task-composition sensitivity ‣ Appendix J Why Occupation Rankings Differ ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">16</a> reports the resulting rank correlations and\nhow often each original leader stays first. Computer &amp; Mathematical retains\nthe NVJ lead in 76.5% of portfolios, while Architecture &amp; Engineering\nretains the TSJ lead in 40.4%. These frequencies describe the specified task removals, not the\nprobability that a category would lead across all professional work.</p>\n<p>Table 16: The NVJ leader survives one-task deletions more often than the TSJ leader. Across 10,000 independently drawn portfolios, one task is chosen uniformly from each category and the remaining tasks are re-averaged within their occupations. GDP occupation weights and recorded outcomes and resource use stay fixed. Percentile ranges describe these perturbations, not confidence intervals.</p>\n<div class=\"table-wrap\"><table><thead><tr><th>Ranking</th><th>Median ρ\\rho</th><th>5th–95th ρ\\rho</th><th>Leader retained</th></tr></thead><tbody><tr><td>Win–tie rate per MJ</td><td>0.951</td><td>0.896–0.989</td><td>51.0%</td></tr><tr><td>NVJ</td><td>0.962</td><td>0.824–0.995</td><td>76.5%</td></tr><tr><td>TSJ</td><td>0.918</td><td>0.813–0.973</td><td>40.4%</td></tr></tbody></table></div>\n<h2 id=\"appendix-k-expert-reference-and-repeated-run-sensitivity\">Appendix K Expert-Reference and Repeated-Run Sensitivity</h2>\n<h3 id=\"k-1-separating-reference-judgments-from-time-valuation\">K.1 Separating reference judgments from time valuation</h3>\n<p>The expert reference affects the metrics through its deliverable, which determines\nwhether the autonomous output wins or ties, and its expert completion time, which determines\nthe output dollar value and expert time assigned to that outcome. We cross these two choices for the same\nGLM 5.2 outputs under the one-run policy, keeping wages, occupation weights,\nAI-system completion time, AI-system execution cost, and energy fixed. The\ncalculation reuses the recorded runs and judge decisions.</p>\n<p>Let s,t∈{E,E+AI}s,t\\in{\\mathrm{E},\\mathrm{E{+}AI}} index the judgment and time\nreferences separately, with E denoting the expert-only reference and E+AI denoting the\nexpert-with-AI reference. Let YisY_{i}^{s} indicate a win or tie against deliverable\nss, HitH_{i}^{t} be expert completion time, WiW_{i} the median occupation\nwage, LiL_{i} the evaluated run’s AI-system completion time, and Ci,EiC_{i},E_{i} its AI-system execution cost and\nenergy. With normalized task weights pi=wo⁡(i)GDP/no⁡(i)p_{i}=w_{o(i)}^{\\mathrm{GDP}}/n_{o(i)},\nwhere no⁡(i)n_{o(i)} is the number of evaluated tasks in occupation o⁡(i)o(i),</p>\n<div class=\"table-wrap\"><table><thead><tr><th></th><th></th><th></th><th></th><th></th></tr></thead><tbody><tr><td></td><td>NVJ⁡(s,t)\\displaystyle\\mathrm{NVJ}(s,t)</td><td>=∑ipi​Yis​Hit​Wi−∑ipi​Ci∑ipi​Ei,\\displaystyle=\\frac{\\sum_{i}p_{i}Y_{i}^{s}H_{i}^{t}W_{i}-\\sum_{i}p_{i}C_{i}}{\\sum_{i}p_{i}E_{i}},</td><td></td><td>(12)</td></tr><tr><td></td><td></td><td></td><td></td><td></td></tr><tr><td>---</td><td>---</td><td>---</td><td>---</td><td>---</td></tr><tr><td></td><td>TSJ⁡(s,t)\\displaystyle\\mathrm{TSJ}(s,t)</td><td>=∑ipi​Yis​(Hit−Li)∑ipi​Ei.\\displaystyle=\\frac{\\sum_{i}p_{i}Y_{i}^{s}(H_{i}^{t}-L_{i})}{\\sum_{i}p_{i}E_{i}}.</td><td></td><td>(13)</td></tr></tbody></table></div>\n<p>Negative time saving is retained. All four combinations charge\n0.429915 MJ and $1.91739 in the GDP-weighted task average.\nTable <a href=\"#A11.T17\" title=\"Table 17 ‣ K.1 Separating reference judgments from time valuation ‣ Appendix K Expert-Reference and Repeated-Run Sensitivity ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">17</a> reports their returns.</p>\n<p>Table 17: Expert-with-AI judgments and completion times each lower reported NVJ and TSJ.\nFor GLM 5.2 under pass​@​1\\mathrm{pass@}1, the mixed-reference rows separate judgment and valuation effects.\nBrackets show percentile 95% task-composition intervals.</p>\n<div class=\"table-wrap\"><table><thead><tr><th>Judgment reference</th><th>Time reference</th><th>NVJ (USD/MJ)</th><th>TSJ (hours/MJ)</th></tr></thead><tbody><tr><td>Expert-only</td><td>Expert-only</td><td>168.11 [126.00, 222.43]</td><td>3.823 [2.870, 5.021]</td></tr><tr><td>Expert-with-AI</td><td>Expert-only</td><td>69.07 [46.04, 95.84]</td><td>1.818 [1.256, 2.486]</td></tr><tr><td>Expert-only</td><td>Expert-with-AI</td><td>84.90 [64.79, 107.49]</td><td>1.990 [1.432, 2.639]</td></tr><tr><td>Expert-with-AI</td><td>Expert-with-AI</td><td>50.04 [30.77, 70.41]</td><td>1.340 [0.814, 1.888]</td></tr></tbody></table></div>\n<h5 id=\"paired-contrasts-and-uncertainty\">Paired contrasts and uncertainty.</h5>\n<p>We resample evaluated tasks with replacement within each of the 44 occupations\n10,000 times (seed 20260726), sharing each draw across all four combinations\nand recomputing ratios and percentage reductions. In a separate sensitivity\nanalysis, we ran each task four times and compared returns using one through\nfour attempts (Appendix <a href=\"#A11.SS2\" title=\"K.2 Repeated runs: Pass@𝑘 and four-run win–tie thresholds ‣ Appendix K Expert-Reference and Repeated-Run Sensitivity ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">K.2</a>). Table <a href=\"#A11.T18\" title=\"Table 18 ‣ Paired contrasts and uncertainty. ‣ K.1 Separating reference judgments from time valuation ‣ Appendix K Expert-Reference and Repeated-Run Sensitivity ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">18</a>\nuses the relevant starting cell as the percentage denominator. The intervals\nvary task composition while holding recorded judgments, valuation inputs,\nand energy allocation fixed.</p>\n<p>Table 18: Changing both judgments and times reduces NVJ by 70.2% and TSJ by 65.0%.\nEvery change substitutes the expert-with-AI reference for the expert-only reference. Values are percentage\nreductions [95% interval]; a negative endpoint permits an increase.</p>\n<div class=\"table-wrap\"><table><thead><tr><th>Change and fixed quantity</th><th>NVJ reduction (%)</th><th>TSJ reduction (%)</th></tr></thead><tbody><tr><td>Judgments; expert-only reference times fixed</td><td>58.9 [43.0, 72.7]</td><td>52.5 [37.0, 66.5]</td></tr><tr><td>Times; expert-only reference decisions fixed</td><td>49.5 [33.1, 63.0]</td><td>47.9 [29.1, 63.2]</td></tr><tr><td>Times; expert-with-AI reference decisions fixed</td><td>27.5 [-1.7, 50.8]</td><td>26.3 [-6.2, 51.8]</td></tr><tr><td>Judgments; expert-with-AI reference times fixed</td><td>41.1 [20.9, 60.4]</td><td>32.7 [7.2, 54.9]</td></tr><tr><td>Both judgments and times</td><td>70.2 [54.4, 82.5]</td><td>65.0 [46.3, 79.4]</td></tr></tbody></table></div>\n<h5 id=\"interaction-and-attribution\">Interaction and attribution.</h5>\n<p>The two reductions from the expert-only reference starting point are not additive.\nSwitching deliverables first gives NVJ 168.11→69.07→50.04168.11\\to 69.07\\to 50.04. The first step removes 58.92 percentage points of the original NVJ and the subsequent time substitution removes 11.32 points, totaling 70.23%. The reverse order gives 168.11→84.90→50.04168.11\\to 84.90\\to 50.04. Averaging the two switching orders assigns 56.7% of the total NVJ decline to reference judgments and 43.3% to time valuation (TSJ: 53.5% and 46.5%).</p>\n<h5 id=\"judged-win-tie-outcomes-and-reference-interpretation\">Judged win–tie outcomes and reference interpretation.</h5>\n<p>The AI system wins or ties less often against the expert-with-AI reference\nthan against the expert-only reference. The assisted deliverables are\ntherefore harder to match on average under this judge, and the two\nreferences can differ in which outputs win or tie. The judgment contrast\njointly captures reference content and judge behavior.</p>\n<h3 id=\"k-2-repeated-runs-pass-kk-and-four-run-win-tie-thresholds\">K.2 Repeated runs: Pass@kk and four-run win–tie thresholds</h3>\n<p>To test how extra attempts affect return, we execute the AI system\nindependently four times per task. Under\npass​@​k\\mathrm{pass@}k, a task earns credit if at least one of kk independent\nattempts wins or ties against the matched expert deliverable.\nThus, pass​@​1\\mathrm{pass@}1 evaluates one attempt per task, whereas\npass​@​4\\mathrm{pass@}4 asks whether any of four attempts wins or ties.\nIncreasing kk adds both another opportunity to produce a win or tie and that run’s\nenergy to the denominator.</p>\n<p>The <em>four-run win–tie threshold</em> is a separate analysis. It always\nevaluates and charges all four runs, but varies how many of those four must win\nor tie. Pass@kk therefore changes the number of opportunities and the energy\ncounted; the four-run win–tie threshold holds both fixed and changes only the\nnumber of wins or ties required.</p>\n<p>Additional runs improve win–tie coverage but reduce return per\njoule. Figure <a href=\"#A11.F10\" title=\"Figure 10 ‣ K.2 Repeated runs: Pass@𝑘 and four-run win–tie thresholds ‣ Appendix K Expert-Reference and Repeated-Run Sensitivity ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">10</a> shows pass​@​1\\mathrm{pass@}1,\npass​@​2\\mathrm{pass@}2, and pass​@​4\\mathrm{pass@}4; NVJ and\nTSJ decline with kk because the increase in tasks with a win or tie does not offset the\nenergy of every additional run. These comparisons charge all scheduled attempts,\nincluding attempts after an earlier output wins or ties; they do not evaluate early\nstopping or show that every retry policy reduces return per joule. Retry\npolicy is therefore an efficiency decision as well as a reliability decision.</p>\n<p>Figure 10: More runs and higher four-run win–tie thresholds both lower measured NVJ and TSJ, but for different reasons. GLM 5.2 uses the expert-only reference. Left: pass@1, pass@2, and pass@4 include the energy of one, two, and four runs per task. Right: the end-to-end energy of four runs is fixed while the four-run win–tie threshold increases. Whiskers are percentile 95% task-composition bootstrap intervals.</p>\n<p>On tasks that match the expert-with-AI reference, the AI system accounts\nfor most of the measured time saving. At the ≥1/4\\geq 1/4 threshold, at least\none of four AI runs wins or ties against the expert-with-AI deliverable. For\nthose tasks, the AI system saves 1.24 GDP-weighted hours per task relative to\nthe expert-only reference. Of this total, 0.43 hours (35.0%) is the difference\nbetween the two expert completion times, and 0.80 hours (65.0%) is the\ndifference between expert-with-AI and AI-system completion times.\nAppendix <a href=\"#A11.SS1\" title=\"K.1 Separating reference judgments from time valuation ‣ Appendix K Expert-Reference and Repeated-Run Sensitivity ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">K.1</a> separates the effects of changing\nthe judged deliverable and completion time while holding AI runs and energy fixed.</p>\n<p>Figure 11: Automation adds time saving beyond expert assistance, but the balance shifts with the success threshold. At lower thresholds, most saved hours come from automation; at stricter thresholds, more come from expert assistance. GLM 5.2 tasks meet each stated win–tie threshold against the expert-with-AI reference. The left panel reports GDP-weighted hours; the right divides each transition’s saved hours by completion time at its starting expert reference. All four runs’ energy is retained. Percentages across transitions use different denominators and should not be added.</p>\n<p>For each task and expert reference, the run order is fixed before evaluating\npass​@​k\\mathrm{pass@}k. Let Yi​rbY_{ir}^{b} indicate a win or tie on run rr,\nRi,kb=∑r=1kYi​rbR_{i,k}^{b}=\\sum_{r=1}^{k}Y_{ir}^{b} count runs that win or tie among the first\nkk, and Qi,kb=𝟏​(Ri,kb≥1)Q_{i,k}^{b}=\\mathbf{1}(R_{i,k}^{b}\\geq 1) indicate whether a task\nhas at least one win or tie. For occupation oo, the exact outcome accounting is</p>\n<div class=\"table-wrap\"><table><thead><tr><th></th><th></th><th></th><th></th><th></th></tr></thead><tbody><tr><td></td><td>Go,kb\\displaystyle G_{o,k}^{b}</td><td>=∑i:o⁡(i)=oQi,kbGib,\\displaystyle=\\sum_{i:o(i)=o}Q_{i,k}^{b}G_{i}^{b},</td><td></td><td>(14)</td></tr><tr><td></td><td>So,kb\\displaystyle S_{o,k}^{b}</td><td>=∑i:o⁡(i)=oQi,kb=1(Hib−∑r=1kYi​rb​Li​rRi,kb),\\displaystyle=\\sum_{\\begin{subarray}{c}i:o(i)=o\\ Q_{i,k}^{b}=1\\end{subarray}}\\left(H_{i}^{b}-\\frac{\\sum_{r=1}^{k}Y_{ir}^{b}L_{ir}}{R_{i,k}^{b}}\\right),</td><td></td><td></td></tr><tr><td></td><td>Eo,k\\displaystyle E_{o,k}</td><td>=∑i:o⁡(i)=o∑r=1kEi​r,\\displaystyle=\\sum_{i:o(i)=o}\\sum_{r=1}^{k}E_{ir},</td><td></td><td></td></tr><tr><td></td><td>Co,k\\displaystyle C_{o,k}</td><td>=∑i:o⁡(i)=o∑r=1kCi​r.\\displaystyle=\\sum_{i:o(i)=o}\\sum_{r=1}^{k}C_{ir}.</td><td></td><td></td></tr></tbody></table></div>\n<p>Output dollar value is counted at most once. Time saving equals expert completion time\nminus mean AI-system completion time among wins and ties when Qi,kb=1Q_{i,k}^{b}=1, and zero otherwise; AI-system\ncompletion times for losses are excluded. Energy and AI-system execution cost include all first kk runs without early\nstopping, regardless of their judgments. This\nbest-of-kk convention is also used by\nAPEX-Agents (<a href=\"#bib.bib4\">Vidgen et al., 2026</a>).</p>\n<p>Table <a href=\"#A11.T19\" title=\"Table 19 ‣ K.2 Repeated runs: Pass@𝑘 and four-run win–tie thresholds ‣ Appendix K Expert-Reference and Repeated-Run Sensitivity ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">19</a> reports all four pass​@​k\\mathrm{pass@}k policies.</p>\n<p>Table 19: Four attempts increase output dollar value but reduce NVJ and TSJ. GLM 5.2 uses the expert-only reference. Every included run incurs cost and energy; output dollar value is counted once. NVJ is USD/MJ and TSJ is hours/MJ. Brackets give conditional 95% task-composition intervals.</p>\n<div class=\"table-wrap\"><table><thead><tr><th>Policy</th><th>Output value (USD)</th><th>AI cost (USD)</th><th>Energy (MJ)</th><th>NVJ [95% CI]</th><th>TSJ [95% CI]</th><th>Net value/ AI cost</th></tr></thead><tbody><tr><td>pass​@​1\\mathrm{pass@}1</td><td>74.19</td><td>1.92</td><td>0.430</td><td>168.1 [126.0, 222.4]</td><td>3.82 [2.87, 5.02]</td><td>37.69</td></tr><tr><td>pass​@​2\\mathrm{pass@}2</td><td>85.75</td><td>4.00</td><td>0.891</td><td>91.8 [71.0, 118.6]</td><td>2.13 [1.65, 2.72]</td><td>20.45</td></tr><tr><td>pass​@​3\\mathrm{pass@}3</td><td>89.41</td><td>5.98</td><td>1.332</td><td>62.7 [48.9, 80.0]</td><td>1.46 [1.16, 1.85]</td><td>13.96</td></tr><tr><td>pass​@​4\\mathrm{pass@}4</td><td>93.42</td><td>7.92</td><td>1.760</td><td>48.6 [38.3, 61.7]</td><td>1.16 [0.93, 1.45]</td><td>10.79</td></tr></tbody></table></div>\n<p>Tables <a href=\"#A11.T20\" title=\"Table 20 ‣ K.2 Repeated runs: Pass@𝑘 and four-run win–tie thresholds ‣ Appendix K Expert-Reference and Repeated-Run Sensitivity ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">20</a> and <a href=\"#A11.T21\" title=\"Table 21 ‣ K.2 Repeated runs: Pass@𝑘 and four-run win–tie thresholds ‣ Appendix K Expert-Reference and Repeated-Run Sensitivity ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">21</a> give complete\nGDP-weighted benchmark outcomes across four-run win–tie thresholds.</p>\n<p>Table 20: Stricter four-run thresholds reduce credited output dollar value and returns at fixed energy. GLM 5.2 uses the expert-only reference. Every included run incurs cost and energy; output dollar value is counted once. NVJ is USD/MJ and TSJ is hours/MJ. Brackets give conditional 95% task-composition intervals.</p>\n<div class=\"table-wrap\"><table><thead><tr><th>Policy</th><th>Output value (USD)</th><th>AI cost (USD)</th><th>Energy (MJ)</th><th>NVJ [95% CI]</th><th>TSJ [95% CI]</th><th>Net value/ AI cost</th></tr></thead><tbody><tr><td>≥1/4\\geq 1/4</td><td>93.42</td><td>7.92</td><td>1.760</td><td>48.6 [38.3, 61.7]</td><td>1.16 [0.93, 1.45]</td><td>10.79</td></tr><tr><td>≥2/4\\geq 2/4</td><td>81.58</td><td>7.92</td><td>1.760</td><td>41.9 [31.7, 54.9]</td><td>1.02 [0.79, 1.31]</td><td>9.30</td></tr><tr><td>≥3/4\\geq 3/4</td><td>71.53</td><td>7.92</td><td>1.760</td><td>36.1 [26.3, 48.8]</td><td>0.89 [0.67, 1.17]</td><td>8.03</td></tr><tr><td>4/44/4</td><td>57.12</td><td>7.92</td><td>1.760</td><td>28.0 [18.4, 40.5]</td><td>0.72 [0.50, 0.99]</td><td>6.21</td></tr></tbody></table></div>\n<p>Against the expert-with-AI reference, requiring a win or tie on at least\nthree of four runs brings output dollar value below AI-system execution cost,\nso NVJ becomes negative at the two strictest thresholds.</p>\n<p>For tasks meeting the same threshold against the expert-with-AI reference, total\nexpert-only reference-to-AI-system time saving decomposes exactly into\nexpert-only reference-to-expert-with-AI reference and\nexpert-with-AI reference-to-AI-system components.</p>\n<p>Table 21: Stricter four-run thresholds turn NVJ negative against the expert-with-AI reference. GLM 5.2 incurs cost and energy for every included run; output dollar value is counted once. NVJ is USD/MJ and TSJ is hours/MJ. Brackets give conditional 95% task-composition intervals.</p>\n<div class=\"table-wrap\"><table><thead><tr><th>Policy</th><th>Output value (USD)</th><th>AI cost (USD)</th><th>Energy (MJ)</th><th>NVJ [95% CI]</th><th>TSJ [95% CI]</th><th>Net value/ AI cost</th></tr></thead><tbody><tr><td>≥1/4\\geq 1/4</td><td>35.11</td><td>7.92</td><td>1.760</td><td>15.4 [10.5, 20.7]</td><td>0.46 [0.32, 0.61]</td><td>3.43</td></tr><tr><td>≥2/4\\geq 2/4</td><td>27.70</td><td>7.92</td><td>1.760</td><td>11.2 [6.1, 16.7]</td><td>0.37 [0.23, 0.52]</td><td>2.50</td></tr><tr><td>≥3/4\\geq 3/4</td><td>7.71</td><td>7.92</td><td>1.760</td><td>-0.1 [-1.9, 2.1]</td><td>0.10 [0.05, 0.16]</td><td>-0.03</td></tr><tr><td>4/44/4</td><td>3.48</td><td>7.92</td><td>1.760</td><td>-2.5 [-3.7, -1.1]</td><td>0.05 [0.01, 0.09]</td><td>-0.56</td></tr></tbody></table></div>\n<p>Table 22: Most measured time saving occurs between the assisted expert reference and autonomous AI output. GDP-weighted task-average decomposition under pass​@​1\\mathrm{pass@}1. Bracketed values are percentile 95% task-composition bootstrap intervals.</p>\n<div class=\"table-wrap\"><table><thead><tr><th>Transition</th><th>Time saving (hours)</th><th>Time saving (%)</th><th>Energy (MJ)</th><th>TSJ (hours/MJ)</th></tr></thead><tbody><tr><td>Expert-only →\\rightarrow expert-with-AI</td><td>0.21 [-0.04, 0.45]</td><td>23.0 [-5.3, 45.9]</td><td>–</td><td>–</td></tr><tr><td>Expert-with-AI →\\rightarrow AI system</td><td>0.58 [0.34, 0.83]</td><td>83.7 [77.8, 87.1]</td><td>0.430</td><td>1.340 [0.814, 1.888]</td></tr><tr><td>Expert-only →\\rightarrow AI system</td><td>0.78 [0.55, 1.03]</td><td>87.4 [84.4, 89.8]</td><td>0.430</td><td>1.818 [1.256, 2.486]</td></tr></tbody></table></div>\n<p>Figure <a href=\"#A11.F11\" title=\"Figure 11 ‣ K.2 Repeated runs: Pass@𝑘 and four-run win–tie thresholds ‣ Appendix K Expert-Reference and Repeated-Run Sensitivity ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">11</a> shows the corresponding four-run threshold\ncomparison; Table <a href=\"#A11.T22\" title=\"Table 22 ‣ K.2 Repeated runs: Pass@𝑘 and four-run win–tie thresholds ‣ Appendix K Expert-Reference and Repeated-Run Sensitivity ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">22</a> gives the pass@1 decomposition.\nThe assisted-expert-to-AI-system component contributes 0.58 hours of the\n0.78-hour expert-only-to-AI-system time saving in that comparison.</p>\n<h2 id=\"appendix-l-five-model-comparison\">Appendix L Five-Model Comparison</h2>\n<p>Table <a href=\"#A12.T23\" title=\"Table 23 ‣ Appendix L Five-Model Comparison ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">23</a> reports GDP-weighted model\noutcomes across the same evaluated tasks and task-composition bootstrap intervals.</p>\n<p>Figure 12: Greater output dollar value need not mean greater economic return per joule. The five models use GDP-weighted aggregates, the expert-only reference, and one attempt per task. NVJ accounts for execution cost as well as energy; TSJ captures expert time saving per unit of energy. Whiskers are percentile 95% task-composition bootstrap intervals.</p>\n<p>Table 23: GLM 5.2 leads output dollar value; Gemma 4 31B leads NVJ and TSJ among five models. The models share the evaluated tasks, the expert-only reference, and a one-attempt policy; values are GDP-weighted task averages. NVJ is USD/MJ, TSJ is hours/MJ, and net value per dollar of AI cost divides net dollar value by AI-system execution cost. Brackets give conditional 95% task-composition intervals. Cost inputs and allocation are in Appendix <a href=\"#A7\" title=\"Appendix G Calculating AI-System Execution Cost ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">G</a>.</p>\n<div class=\"table-wrap\"><table><thead><tr><th>Model</th><th>GPUs</th><th>Output value (USD)</th><th>AI cost (USD)</th><th>Energy (MJ)</th><th>NVJ [95% CI]</th><th>TSJ [95% CI]</th><th>Net value/ AI cost</th></tr></thead><tbody><tr><td>GLM 5.2</td><td>8×\\timesH200</td><td>74.19</td><td>1.92</td><td>0.430</td><td>168.1 [126.0, 222.4]</td><td>3.82 [2.87, 5.02]</td><td>37.69</td></tr><tr><td>Qwen3.6-27B</td><td>8×\\timesH100</td><td>67.81</td><td>4.12</td><td>0.516</td><td>123.4 [92.2, 163.2]</td><td>2.43 [1.73, 3.28]</td><td>15.46</td></tr><tr><td>Gemma 4 31B</td><td>8×\\timesH100</td><td>41.38</td><td>0.86</td><td>0.097</td><td>417.5 [239.5, 642.9]</td><td>10.86 [6.76, 15.72]</td><td>47.38</td></tr><tr><td>GPT-OSS 120B</td><td>8×\\timesH100</td><td>39.73</td><td>1.98</td><td>0.137</td><td>275.7 [152.5, 429.2]</td><td>6.38 [3.71, 9.63]</td><td>19.09</td></tr><tr><td>GPT-OSS 20B</td><td>8×\\timesH100</td><td>35.88</td><td>1.12</td><td>0.103</td><td>336.0 [180.7, 536.4]</td><td>8.26 [4.67, 12.59]</td><td>31.03</td></tr></tbody></table></div>\n<p>Figure <a href=\"#A12.F12\" title=\"Figure 12 ‣ Appendix L Five-Model Comparison ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">12</a> summarizes output dollar value, NVJ, and\nTSJ, while Table <a href=\"#A12.T23\" title=\"Table 23 ‣ Appendix L Five-Model Comparison ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">23</a> also provides the energy denominator,\nGDP-weighted task averages, and intervals.</p>\n<p>Occupation-category rankings are broadly similar across models,\nespecially for output dollar value and time saving. NVJ and TSJ rankings vary\nmore because the evaluated models have different end-to-end energy profiles. Table <a href=\"#A12.T24\" title=\"Table 24 ‣ Appendix L Five-Model Comparison ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">24</a>\nreports all pairwise rank correlations.</p>\n<p>Table 24: Category output-dollar-value rankings agree across models more than energy-return rankings do on average. Spearman correlations across the 13 categories after including AI-system execution cost in NVJ.</p>\n<p>Model A\nModel B\nOutput dollar value\nTime saving\nNVJ\nTSJ</p>\n<p>GLM 5.2\nQwen3.6-27B\n0.830\n0.703\n0.780\n0.544</p>\n<p>GLM 5.2\nGemma 4 31B\n0.654\n0.912\n0.560\n0.731</p>\n<p>GLM 5.2\nGPT-OSS 120B\n0.637\n0.742\n0.478\n0.478</p>\n<p>GLM 5.2\nGPT-OSS 20B\n0.643\n0.824\n0.489\n0.626</p>\n<p>Qwen3.6-27B\nGemma 4 31B\n0.824\n0.665\n0.549\n0.302</p>\n<p>Qwen3.6-27B\nGPT-OSS 120B\n0.769\n0.659\n0.643\n0.692</p>\n<p>Qwen3.6-27B\nGPT-OSS 20B\n0.808\n0.885\n0.560\n0.659</p>\n<p>Gemma 4 31B\nGPT-OSS 120B\n0.802\n0.808\n0.863\n0.654</p>\n<p>Gemma 4 31B\nGPT-OSS 20B\n0.907\n0.802\n0.742\n0.692</p>\n<p>GPT-OSS 120B\nGPT-OSS 20B\n0.863\n0.863\n0.879\n0.846</p>\n<h2 id=\"appendix-m-accelerator-comparisons\">Appendix M Accelerator Comparisons</h2>\n<p>Table <a href=\"#A13.T25\" title=\"Table 25 ‣ Appendix M Accelerator Comparisons ‣ Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?\">25</a> reports the resulting GDP-weighted\noutcomes and economic return per joule.</p>\n<p>Table 25: B200 has the highest NVJ and TSJ among the four single accelerators. Qwen3.6-27B is evaluated against the expert-only reference under the paper’s GDP-based weights. Each accelerator is summarized from its own recorded outcomes and measured energy. Brackets give 95% task-composition intervals; costs and energy include every attempt.</p>\n<p>Accelerator\nNVJ (USD/MJ)\nTSJ (hours/MJ)</p>\n<p>NVIDIA B200\n535.5 [388.4, 721.6]\n11.684 [8.416, 15.620]</p>\n<p>NVIDIA H200\n391.3 [274.9, 532.4]\n7.293 [4.864, 10.188]</p>\n<p>NVIDIA H100\n402.2 [298.6, 531.7]\n7.243 [5.122, 9.934]</p>\n<p>AMD MI350X\n124.1 [81.7, 177.8]\n1.260 [0.450, 2.297]</p>\n<p>One and eight H100s use the same evaluated tasks, expert-only reference,\npass​@​1\\mathrm{pass@}1 rule, and GDP weights, with separately recorded outcomes.\nTheir respective output dollar values, execution costs, and energies are\n($64.79,$0.71,0.159​MJ)(\\$64.79,\\$0.71,0.159\\,\\mathrm{MJ}) and\n($67.81,$4.12,0.516​MJ)(\\$67.81,\\$4.12,0.516\\,\\mathrm{MJ}), yielding NVJ of 402.16 versus\n123.39 USD/MJ and TSJ of 7.243 versus 2.432 hours/MJ.</p>\n<h2 id=\"appendix-n-gdpworks-83-additional-occupational-tasks\">Appendix N GDPWorks: 83 Additional Occupational Tasks</h2>\n<p>To broaden professional-work evaluation beyond GDPval, we assembled\n<em>GDPWorks</em>, a dataset of 83 additional tasks created by Snorkel AI. Each task includes\na request, grading rubric, source files, expert-only and expert-with-AI\ndeliverables, and time annotations. Together with OpenAI’s 220 public GDPval\ntasks and the annotations added by Snorkel AI, these materials are prepared\nfor release by the Scaling Intelligence Lab and HazyResearch at\n<a href=\"https://huggingface.co/datasets/ScalingIntelligence/joules-to-gdp\" rel=\"nofollow ugc noopener\">ScalingIntelligence/joules-to-gdp</a>,\nwith GDPWorks in the joules_to_gdp split.</p>\n<p>GDPWorks adds 36 detailed O*NET occupations absent from GDPval’s 44,\nbringing the combined coverage to 80 distinct occupation codes. Examples\ninclude bioinformatics scientists, environmental scientists, and purchasing\nmanagers. The tasks introduce recent work contexts and 2026-dated source\nmaterials, including financial records and environmental monitoring data.\nThey also target more demanding, source-intensive workflows: median supplied\nexpert-only time estimates are 4.0 hours, compared with 2.34 hours for GDPval,\nand the median number of included reference files is six, compared with one.\nThese time estimates characterize expected expert effort. GDPWorks\nsupports future tests of how economic return per joule generalizes to\nadditional occupations and longer workflows. GDPWorks is not a\nGDP-weighted sample; the results reported in this paper use the 220-task\nGDPval split.</p>","headings":[{"level":1,"text":"Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?","id":"joules-to-gdp-how-much-economic-value-do-ai-systems-produce-per-"},{"level":2,"text":"1 Introduction","id":"1-introduction"},{"level":2,"text":"2 Methods","id":"2-methods"},{"level":3,"text":"2.1 Dataset and expert references","id":"2-1-dataset-and-expert-references"},{"level":3,"text":"2.2 Blinded evaluation","id":"2-2-blinded-evaluation"},{"level":3,"text":"2.3 End-to-end energy accounting","id":"2-3-end-to-end-energy-accounting"},{"level":3,"text":"2.4 Net Value per Joule and Time Saving per Joule","id":"2-4-net-value-per-joule-and-time-saving-per-joule"},{"level":3,"text":"2.5 Two levels of GDP-based aggregation","id":"2-5-two-levels-of-gdp-based-aggregation"},{"level":2,"text":"3 Experiments and Results","id":"3-experiments-and-results"},{"level":3,"text":"3.1 How does economic return per joule vary across occupations? (Q1)","id":"3-1-how-does-economic-return-per-joule-vary-across-occupations-q"},{"level":3,"text":"3.2 How do models trade off economic value and energy? (Q2)","id":"3-2-how-do-models-trade-off-economic-value-and-energy-q2"},{"level":3,"text":"3.3 How do accelerator type and count affect return per joule? (Q3)","id":"3-3-how-do-accelerator-type-and-count-affect-return-per-joule-q3"},{"level":2,"text":"4 Limitations and Future Work","id":"4-limitations-and-future-work"},{"level":2,"text":"5 Conclusion","id":"5-conclusion"},{"level":2,"text":"Reproducibility Statement","id":"reproducibility-statement"},{"level":2,"text":"Acknowledgements","id":"acknowledgements"},{"level":2,"text":"References","id":"references"},{"level":2,"text":"Appendix A Related Work","id":"appendix-a-related-work"},{"level":2,"text":"Appendix B Expert References and Annotations","id":"appendix-b-expert-references-and-annotations"},{"level":2,"text":"Appendix C Benchmark Contamination","id":"appendix-c-benchmark-contamination"},{"level":2,"text":"Appendix D Judge and Evidence Pipeline","id":"appendix-d-judge-and-evidence-pipeline"},{"level":3,"text":"D.1 Judge consistency","id":"d-1-judge-consistency"},{"level":2,"text":"Appendix E Execution and End-to-End Energy Accounting","id":"appendix-e-execution-and-end-to-end-energy-accounting"},{"level":2,"text":"Appendix F Energy-Attribution Sensitivity","id":"appendix-f-energy-attribution-sensitivity"},{"level":2,"text":"Appendix G Calculating AI-System Execution Cost","id":"appendix-g-calculating-ai-system-execution-cost"},{"level":3,"text":"G.1 Execution-cost sensitivity","id":"g-1-execution-cost-sensitivity"},{"level":2,"text":"Appendix H Economic Weights and Wage Robustness","id":"appendix-h-economic-weights-and-wage-robustness"},{"level":2,"text":"Appendix I Occupation-Category Results","id":"appendix-i-occupation-category-results"},{"level":2,"text":"Appendix J Why Occupation Rankings Differ","id":"appendix-j-why-occupation-rankings-differ"},{"level":3,"text":"J.1 Full Occupation-Category Results","id":"j-1-full-occupation-category-results"},{"level":3,"text":"J.2 Task-composition sensitivity","id":"j-2-task-composition-sensitivity"},{"level":2,"text":"Appendix K Expert-Reference and Repeated-Run Sensitivity","id":"appendix-k-expert-reference-and-repeated-run-sensitivity"},{"level":3,"text":"K.1 Separating reference judgments from time valuation","id":"k-1-separating-reference-judgments-from-time-valuation"},{"level":3,"text":"K.2 Repeated runs: Pass@kk and four-run win–tie thresholds","id":"k-2-repeated-runs-pass-kk-and-four-run-win-tie-thresholds"},{"level":2,"text":"Appendix L Five-Model Comparison","id":"appendix-l-five-model-comparison"},{"level":2,"text":"Appendix M Accelerator Comparisons","id":"appendix-m-accelerator-comparisons"},{"level":2,"text":"Appendix N GDPWorks: 83 Additional Occupational Tasks","id":"appendix-n-gdpworks-83-additional-occupational-tasks"}]}}