How digital assets connect intelligence, commerce, and compute
\nBlackRock Digital Assets Research
\nThe Machine-\nNative Economy\nHow digital assets connect intelligence,\ncommerce, and compute
\nExecutive Summary
\nThe extraordinary growth in artificial intelligence is the defining technology theme of this era. The rise of\ndigital assets, meanwhile, represents a concurrent technology theme with particularly relevant implications\nfor financial infrastructure. These themes have historically developed along largely parallel tracks, but they\nare beginning to converge as AI systems become more capable of interacting with financial and economic\nnetworks. This paper examines this growing relationship and explains why broad AI adoption may represent\nan underappreciated source of demand, utility, and application growth across the digital asset economy.
\nAt the core of this convergence, AI and digital assets both arise from a common foundation: AI represents\nmachine-native intelligence, while digital assets represent machine-native money. This alignment\nbecomes particularly important with the rise of agentic AI, which refers to systems that can plan and\nexecute multistep tasks toward a defined objective by interacting with external tools and infrastructure with\nlimited human intervention, with blockchains providing the programmable infrastructure that connects\nintelligence with economic activity. These capabilities extend AI beyond content generation toward real-\nworld action, including making purchases and initiating financial transactions.
\nSpecifically, we explore three key areas of overlap:
\n• LLMs and blockchains share analogous tokenization architectures. Large language models divide\n human language into tokens and encode them numerically for model interpretation and processing.\n Blockchains similarly represent economic value and entitlements as digital asset tokens designed for\n machine-verifiable transfer and settlement. The functions are distinct, but both workflows translate real-\n world inputs into formats that machines can use natively.
\n• Agentic commerce requires machine-native payment rails. The rise of agentic AI and machine-to-\n machine payments will likely increase demand for blockchains and other programmable payment\n infrastructure; stablecoins, native cryptoassets, and other on-chain assets can serve as machine-native\n instruments for payment and settlement across these rails. Existing rails such as ACH and card\n networks support substantial automation, although their onboarding requirements and settlement\n economics can make them less suited to always-on, very low-value transactions requiring\n programmable execution. Emerging protocols such as x402 and ACP are being deployed on blockchain\n networks and alongside adaptations to traditional payment rails, creating a transaction and settlement\n layer for more complex agentic workflows.
\n• Compute is emerging as a new and potentially large market for digital assets. Compute, the\n processing capacity required to train and run AI systems, is becoming an increasingly important\n economic resource. Analyst estimates suggest that hyperscaler cloud revenues could exceed $1 trillion\n annually by 2030. As agents become more capable and persistent, standardized claims on compute\n capacity could become a significant digital asset use case for financing and programmable settlement.
\nTogether, these developments position AI as a structural catalyst for digital asset adoption and digital\nassets as a potential facilitator of the AI economy: AI interprets information and directs action, while\nblockchains provide machine-readable assets and programmable settlement. This relationship remains\nunderappreciated and could expand the role of digital assets as core infrastructure for an increasingly\nautonomous digital economy.
\nAI and digital asset tokenization convert real-\nworld inputs into machine-native representations\nAt the architectural level, tokenization in artificial intelligence and digital assets serves an analogous\npurpose: translating information and economic entitlements into discrete, standardized representations\nthat machines can process natively. AI tokens encode information, while digital asset tokens represent\nunits of value, ownership, or economic claims. In large language models (LLMs), a tokenizer divides\nhuman-readable text into smaller units, typically words, sub-words, or characters, and maps them to\nnumerical identifiers. An embedding layer then converts those identifiers into vector representations that\nthe model can use for computation. The model generates numerical token IDs that are decoded back into\nhuman-readable text. By continuously translating incoming text into standardized numerical units, LLMs\ncan process large input streams efficiently using the highly parallel calculations for which modern chips\nare optimized.
\nBlockchains apply a functionally comparable, though technically distinct, process to stores of value and\neconomic claims. Digital asset tokenization is the process of representing a financial or real-world asset\n(RWA) such as cash, a fund interest, a security, or another ownership claim as a standardized digital token\nrecorded on a distributed ledger. Existing on-chain assets such as stablecoins can be transferred through\nsmart contract transactions. At the transaction level, these assets are expressed through machine-readable\ntransaction data and governing rules. The network verifies authorization and transaction validity, while\nsmart contracts or transaction scripts apply asset-specific permissions and conditions. Together, these\nmechanisms update the ledger by transferring token balances. Once a transaction has been recorded and\nsufficiently finalized under the network’s consensus rules, it becomes part of the network’s canonical\nledger state.
\nIn the blockchain workflow, smart contracts execute transaction logic and apply rules specific to financial\nassets. Anti-money laundering (AML), know-your-customer (KYC), and know-your-agent (KYA) checks\ngenerally occur off-chain, where identity and compliance data can be assessed, with verified results passed\non-chain to determine transaction eligibility. For tokenized RWAs, these controls sit within a broader legal\nand regulated-service framework that relies on authoritative off-chain registries, while reducing\ndependence on closed databases and manual reconciliation.
\n 3Figure 1: Illustrative Tokenization Workflows in LLMs and Blockchain Transactions
\n AI Tokenization: DA Tokenization:\n Text to LLM Representation RWAs to On-Chain Ownership Record\n\n Text Real-world asset (RWA)\n\n $100 beneficial interest in a\n “AI is changing the world”\n money market fund\n\n Tokenized Digital asset\n AI tokenization\n Tokens ownership tokenization\n\n Tokenized fund shares assigned\n “AI”, “is”, “changing”, “the”, “world”\n to investor’s digital wallet\n\n Encoded Validation, locking,\n Encoding\n Numerical IDs transaction fields encoding\n\n [Asset ID, Sender ID, Recipient ID,\n [101, 23, 4587, 5, 982]\n Units, Eligibility Flag]\n\n\n Execution and\n Embedding Embedding Blockchain finalization\n matrix transaction record\n\n tx_hash: 0x8F3A...91C2\n [0.48, -0.37, 0.20, -0.88, 0.12] asset_id: 0x2D7B...44E1\n [-0.62, 0.88, -0.45, 0.32, 0.38] from_id: 0x91C3...0A77\n [0.29, -0.55, 0.90, -0.62, 0.45] to_id: 0x71A9...3F06\n [-0.07, 0.32, 0.79, 0.50, -0.92] token_units: 100.000000\n [-0.34, 0.15, 0.68, 0.72, -0.22] block: 21845902\n status: finalizedFor illustrative purposes only. Source: BlackRock Digital Assets Research.
\nThese parallel workflows, one encoding human context and the other encoding economic entitlement, are\nbecoming more relevant as systems of intelligence and transaction execution converge. Both systems use\nstructured, machine-readable representations, which can give LLM-based AI agents a more direct\ninterface with blockchain data than with many fragmented legacy systems. In turn, greater standardization\nacross asset classes can reduce reliance on bespoke integrations and make it easier for agents to\norchestrate more complex multi-asset workflows. Through programmable interfaces, agents can evaluate\nbalances and rules before executing authorized transactions and verifying settlement, with limited reliance\non manual processes.
\nRecent Bitcoin Policy Institute research offers preliminary support for this framework, reporting that model\noutputs across controlled simulations generally favored stablecoins for everyday payments and bitcoin for\nlong-term value preservation.1 These findings reflect simulated model responses rather than observed\nagent behavior, but point to a potential AI-native monetary architecture in which stablecoins serve as\ntransaction money and bitcoin as a store of value. As we explore in the next section, this shared machine-\nnative foundation could support meaningful agentic payment use cases that some traditional financial\nrails may serve less efficiently or economically.
\nBitcoin Policy Institute, “Which money do AI agents prefer?”, https://www.btcpolicy.org/articles/study-ai-models-overwhelmingly-prefer-bitcoin-and-digital-native-money-over-
traditional-fiat.\n 4
Machine-native transactions require purpose-\nbuild agentic payment protocols\nAs AI agents become more capable and as their real-world applications expand, they increasingly demand\npayment and asset infrastructure designed natively for machine-speed commerce. Crypto-native\nblockchain rails are particularly well suited to high-frequency, sub-cent, machine-to-machine (M2M)\ntransactions that take place around-the-clock, including API calls, on-demand data, and consumption-\nbased compute. In parallel, modified traditional payment systems will remain important for connecting\nagents with human-operated businesses and consumers in business-to-machine (B2M) and consumer-to-\nmachine (C2M) settings.
\nStablecoins, native cryptoassets, tokenized real-world assets, and other programmable instruments can\nsupport transactions and digital ownership, including collateral use, with the required granularity on an\nalways-on basis. More broadly, tokenization can provide standardized, machine-readable representations\nacross asset classes, reducing bespoke integrations across financial infrastructure and enabling agents to\norchestrate increasingly complex workflows more efficiently.
\nMany existing payment rails are less well suited to high-volume, low-denomination agentic transactions.\nThey involve:\n • Account setup, credentialing, and authorization processes that may require human involvement;\n • Merchant acceptance fees that can make very low-value transactions uneconomic;\n • Settlement and finality constraints, as most ACH volume settles within one business day or less,\n while card authorization is near-instant but merchant settlement and dispute finality can take\n longer; and\n • Potential scalability limitations as machine-generated transaction volumes grow.
\nAgentic payment protocols build on foundational standards such as MCP and A2A, which connect agents\nwith external systems and one another. MCP2 (Model Context Protocol), introduced by Anthropic in\nNovember 2024, standardizes how AI applications access external data and workflows. A2A3\n(Agent2Agent), launched by Google in April 2025, enables agents to communicate and coordinate across\nplatforms.
\nAgents using MCP and A2A leverage multiple agentic payment protocols to complete complex workflows\ninvolving payments. x4024, an open payment protocol developed by Coinbase, uses the HTTP 402\n“Payment Required” status code to facilitate machine-initiated payments. The protocol is blockchain-\nagnostic, with stablecoins such as USDC representing an early primary use case, and is emerging as one\npotential standard for high-velocity M2M transactions. By providing 24/7, near-real-time, verifiable\nsettlement, x402 can reduce the resource provider’s counterparty exposure and enable the immediate\nrelease of requested data or services upon payment confirmation. Because x402 uses digital currencies,\nincluding stablecoins held in on-chain wallets, it can support high-frequency, low-denomination\ntransactions without human intervention. Where settlement occurs on permissionless networks, greater\nusage could increase demand for blockspace and validator services, creating a potential transmission\nchannel to native cryptoassets. The extent of value capture will depend on each network’s fee, staking, and\ngas-sponsorship design.
\nModel Context Protocol, https://modelcontextprotocol.io/docs/getting-started/intro. Analysis based on the currency being held as cash. 3. Google Codelabs, https://codelabs.
developers.google.com/intro-a2a-purchasing-concierge#0. 4. Coinbase, x402, https://x402.org/wp-content/uploads/sites/10/2026/06/x402-whitepaper.pdf.\n 5
Other emerging standards connect agentic transactions with existing payment rails and establish\nguardrails for trusted financial execution. The Machine Payments Protocol (MPP5), developed by Stripe\nand Tempo, enables payments for APIs and other HTTP resources with flexible settlement via stablecoins\nor traditional payment methods. The Agentic Commerce Protocol (ACP6), developed by Stripe and OpenAI,\nenables programmatic checkout between agents and businesses while allowing sellers to retain their\nexisting commerce and payment infrastructure. Google’s Agents Payment Protocol (AP27) uses\ncryptographic mandates and audit trails to provide evidence of user authorization, while Visa’s Trusted\nAgents Protocol (TAP8) helps merchants verify trusted agents and securely receive payment credentials.
\nFigure 2 illustrates a simplified agentic payment workflow. A human user (1) asks an agent to book flights\nand hotels within a specified budget. The primary AI agent (2) accesses the user’s calendar, preferences,\nand approved payment details through MCP connectors, then (3) delegates data-gathering to a specialized\nAI travel sub-agent through A2A. The sub-agent (4) calls paid airfare and hotel-rate APIs, with applicable\npayments handled through x402 and settled on-chain. Using the returned information, the primary agent\n(5) completes reservations through airline and hotel checkout systems and (6) returns the itinerary and\nreceipts to the user.
\nFigure 2: Illustrative Agentic Workflow Involving Foundational and Financial Protocols
\n 5MCP: helps agents access external tools and data\nA2A: helps agents communicate with each other ACP\nx402: powers fast machine-to-machine transactions\nACP: connects agents to vendors' existing payment rails Finalize\n Reservations Airline/Hotel\n Checkout
\n 1 3 4\n\n User prompt A2A x402\n\n "Book my trip for "Get schedules Agent pays for data\n under $2,500" Primary and fares" AI Travel to build optimal Airfare/Room\nHuman\n User Agent Sub-Agent itinerary Rates API\n Itinerary, Receipts Data driving\n purchase decisions\n 6 2 MCP MCP\n\n\n\n Tools/Data Tools/Data\n\n\n\n\n Calendar, Email, Payment Credentials Routing, Seat Maps, LoyaltyFor illustrative purposes only. Source: BlackRock Digital Assets Research.
\nSeveral types of digital assets may support agentic commerce, but stablecoins are likely to lead\ntransactional use. Stablecoins are digital tokens designed to maintain a stable value relative to a reference\ncurrency, most commonly the U.S. dollar. Their price stability provides a reliable unit of account and greater\npredictability in pricing and settlement. Stablecoins represent the largest category of tokenized real-world\nassets, with more than $300 billion in circulating market capitalization as of September 2026.9 Adjusted\nstablecoin transaction volume exceeded $11 trillion in 2025, placing it in the same broad range as Visa and\nMastercard’s annual payment volumes.10 Adjusted stablecoin volume remained well below the $93 trillion\ntransferred over ACH in 2025; from 2020 to 2025, however, it grew at an 80% CAGR, compared with\napproximately 8.5% for ACH.11 Growing regulatory clarity, including the GENIUS Act in the U.S., MiCA in the\nEU, Hong Kong’s stablecoin licensing regime, and Singapore’s stablecoin regulatory framework, should\nsupport continued stablecoin adoption and growth.
\nTempo & Stripe, Machine Payments Protocol, https://mpp.dev/overview. 6. Stripe, Agentic Commerce Protocol, https://www.agenticcommerce.dev/docs. 7. Google, Agentic
Payments Protocol, https://ap2-protocol.org/. 8. Visa, Trusted Agent Protocol, https://developer.visa.com/capabilities/trusted-agent-protocol. 9. RWA.xyz, https://app.rwa.xyz/\nstablecoins. 10. Visa/Allium, https://visaonchainanalytics.com/transactions. Adjusted stablecoin transaction volume uses address labels and heuristic filters to exclude internal\ntransfers, intra-exchange flows, bots, and other high-frequency or high-volume activity. 11. Nacha, https://www.nacha.org/content/ach-network-volume-and-value-statistics\n 6
Figure 3: Stablecoin and Major Card Networks Transaction Volumes\n Stablecoins Visa Mastercard\n 20\n 16.7\n 15\nUS $ Trillions
\n 11.2\n 10 10.6 8.5\n\n 5\n\n\n 0\n 2018 2019 2020 2021 2022 2023 2024 2025 1H20262026 stablecoin data through June 2026. Note: measures are not directly comparable. Adjusted stablecoin volume includes selected exchange, DeFi, lending, mint/burn, and ramp\nactivity after methodological filters. Visa reports total volume, including payment and cash volume, while Mastercard reports gross dollar volume. Visa data are for the fiscal year\nended Sept. 30, 2025; Mastercard and stablecoin data are presented on a calendar-year basis. Source: Visa Onchain Analytics, Allium, Visa and Mastercard annual reports, and SEC\nfilings.
\nFor digital assets, the implications of this growth extend beyond stablecoins to the blockchain networks on\nwhich they are issued and settled. Many major stablecoins are issued across multiple blockchains, allowing\nmarket participants to select among supported settlement venues based on economics and technical fit.\nThese venues include general-purpose permissionless networks such as Ethereum as well as purpose-built\nstablecoin networks such as Circle’s Arc, where USDC is designed to serve as the native gas asset. On\npermissionless networks, native cryptoassets (e.g. ETH) support consensus, validator compensation,\ntransaction fees, and settlement. As stablecoin activity scales, greater demand for blockspace and network\nservices could support usage-related demand and potential value capture for these assets, subject to each\nnetwork’s fee, staking, and gas-sponsorship design. Arc also offers a complementary model in which\ngreater payment activity could deepen USDC’s utility as both a settlement asset and the means of paying\ntransaction fees.
\nCompute is emerging as a new and potentially\nlarge market for digital assets as autonomous\nagents proliferate\nAI systems and agents require substantial computing power and energy to operate. Investors have largely\nfocused on the sheer scale of capital expenditures (CapEx) needed to build AI infrastructure, with some\nestimates placing cumulative AI capital spending above $5 trillion between 2025 and 2030. 12 But the\naccompanying rise in ongoing operating expenses (OpEx) to support AI deployments deserves equal\nattention. A meaningful share of this operating spend flows through the AI cloud compute market, which\nmonetizes access to installed IT equipment and the electricity required to operate it. As this market\nexpands, compute is becoming a distinct, large, and increasingly investable economic resource that could\nsupport a new class of digital assets. Using hyperscalers’ major cloud segments as a broad proxy for\nmarket scale, consensus estimates for Amazon Web Services (AWS), Microsoft’s Intelligent Cloud segment,\nand Google Cloud imply combined revenue of approximately $1.1 trillion by 2030, representing a 29%\nCAGR from 2025 levels.13
\nGoldman Sachs, “Private markets are expected to have a growing role in data center financing,” https://www.goldmansachs.com/insights/articles/private-markets-expected-to-
have-growing-role-in-data-center-financing. 13. Bloomberg-compiled sell-side analyst estimates as of Aug. 31, 2026.\n 7
As compute becomes a larger economic input, the need to price and allocate capacity while supporting\nfinancing and hedging should also increase. Historically, large resource markets have developed trading\ninfrastructure that improves liquidity and risk management. In our view, compute may follow a similar path\nas AI adoption scales. Recent market developments already point in this direction, including GPU-backed\nfinancings and financing platforms designed around long-duration, usage-linked compute revenue. These\nstructures reflect the capital intensity of securing leading-edge GPUs and building capacity for frontier-\nmodel development. Demand for both training and inference should continue to grow as models improve.\nAs real-world AI use cases expand, inference is expected to become the largest AI workload by 2030 and\naccount for a growing share of data-center power demand.14 The potential inference user base, spanning\nenterprises and individual consumers, is considerably larger and more fragmented than the concentrated\nset of training-market participants.
\nFigure 4: Global Data Center Power Demand by Workload\n AI Training AI Inference Non-AI
\n 250\n\n\n 200Gigawatts (GW)
\n 150\n\n\n 43%\n 100\n\n\n 50\n 25% 28%\n 28%\n 0\n 2025 2026E 2027E 2028E 2029E 2030EPercentages show each category’s share of total data center power demand. Source: McKinsey Data Center Demand Model, estimates and projections as of Dec. 2025.
\nEven as this market takes shape, meaningful contract-design and market-structure challenges remain\nbefore standardized compute products can scale. These include accounting for substantial differences in\nchip-generation productivity and regional economics, particularly where energy costs diverge, as well as\nestablishing workable standards for both cash settlement and delivery of contracted capacity. We view\nthese as important but ultimately resolvable design considerations. The development of basis markets,\ncontracts for difference, and other mechanisms used in established commodity markets offers a useful\nprecedent for managing heterogeneous assets and localized pricing, while on-chain tokenized markets\nmay enable more granular regional and hardware-specific contracts within shared settlement\ninfrastructure. As these frameworks mature, we expect standardized products, including exchange-traded\ncompute futures, to support more transparent price discovery and more effective hedging for both\nproviders and consumers of compute capacity. Standardized compute contracts could also create claims\non compute capacity and related usage rights that can be represented, transferred, pledged as collateral,\nand settled through programmable infrastructure. This could in turn broaden institutional investor\nparticipation and establish compute as a new opportunity for the broader digital asset ecosystem.
\nMcKinsey, “The next big shifts in AI workloads and hyperscaler strategies,” https://www.mckinsey.com/industries/technology-media-and-telecommunications/our-
insights/the-next-big-shifts-in-ai-workloads-and-hyperscaler-strategies.\n 8
This functionality may become particularly valuable as agentic AI adoption scales and agents increasingly\ndiscover, provision, optimize, and pay for compute through programmable payment rails such as x402.\nAgents could query real-time marketplace APIs to compare available capacity by price, performance,\nlatency, location, and hardware specialization, then provision the resources best suited to a given workload.\nMCP and A2A could facilitate data access and agent-to-agent coordination, while x402 could support on-\ndemand settlement on a per-use, per-model-token, or per-job basis. This framework could enable elastic,\njust-in-time access to compute with limited human intervention. Although agentic payment activity\nremains nascent today, the structural fit between autonomous agents and machine-native payments\nmakes this an area worth monitoring as the ecosystem develops.
\nFigure 5: Illustrative Agentic Workflow for Optimizing and Securing On-Demand Compute\nResources
\n Compute Providers\n\n\n\n 1\n User prompt\n GPU/CPU providers, specialized\n "Run extended compute, edge/regional nodes\n Human analysis" Agent\n User\n Returns output\n\n MCP Discovery\n\n\n\n Tools/Data\n\n Real-time pricing, availability,\n and reliability metrics\n\n Continously executes task and\n estimates compute needs\n Compute capacity delivered to\n support agentic taskFor illustrative purposes only. Source: BlackRock Digital Assets Research.
\nStripe’s August 2026 agreement to acquire OpenRouter provides an early strategic signal that model\nrouting and compute-usage optimization are becoming part of the financial infrastructure surrounding\nAI.15 OpenRouter distributes workloads across more than 400 models from over 80 providers based on\nworkload needs and cost-performance tradeoffs, highlighting the economic value of allocating scarce\ncompute efficiently. The buyer also matters to the thesis: given Stripe’s broader activity across payments,\nstablecoins, billing, and agentic commerce, this transaction points to a potential convergence between\ncompute procurement, usage-based billing, and programmable settlement. In our view, this could support\na future in which agents autonomously source and pay for compute over blockchains and other\nprogrammable payment rails.
\nStripe, “Stripe agrees to acquire OpenRouter to help businesses optimize token routing and usage,” https://stripe.com/newsroom/news/stripe-agrees-to-acquire-openrouter.
9Conclusion\n Authors
\nAI and blockchain-based digital assets are Will Su\nincreasingly converging as machines take a greater Head of Digital Assets\nrole in economic activity. The structured, machine- Research\nreadable representations created through LLM and\nblockchain tokenization can give AI agents a more\ndirect interface with programmable assets, while Robert Mitchnick\nstablecoins and protocols such as x402 may support\n Head of Digital Assets\nhigh-frequency, low-value, always-on transactions. At\nthe same time, standardized and liquid markets for\ncompute claims could allow agents to source,\noptimize, finance, and pay for computing resources as Jay Jacobs\ninference demand expands. The ecosystem remains\n U.S. Head of Equity\nnascent, with agentic payment activity and compute-\n ETFs\nmarket liquidity still limited. As AI adoption broadens\nand agentic systems become more capable, digital\nassets could become increasingly integral to AI’s William Helm\neconomic infrastructure, expanding utility across Head of U.S. iShares\nstablecoins, tokenized RWAs, and native cryptoassets Product Innovation\nthat support blockchain settlement.
\n 10This material is not intended to be relied upon as a forecast, research or investment advice, and is not a recommendation, offer or solicitation to\nbuy or sell any securities or to adopt any investment strategy. The opinions expressed are as of the date indicated and may change as\nsubsequent conditions vary. The information and opinions contained in this material are derived from proprietary and nonproprietary sources\ndeemed by BlackRock to be reliable, are not necessarily all-inclusive and are not guaranteed as to accuracy. As such, no warranty of accuracy or\nreliability is given and no responsibility arising in any other way for errors and omissions (including responsibility to any person by reason of\nnegligence) is accepted by BlackRock, its officers, employees or agents. This material may contain “forward-looking” information that is not\npurely historical in nature. Such information may include, among other things, projections and forecasts. There is no guarantee that any of these\nviews will come to pass. Reliance upon information in this material is at the sole discretion of the viewer.
\nThis material is not intended as an offer or solicitation for the purchase or sale of any security. Specific companies or issuers are mentioned for\neducational purposes only and should not be deemed as a recommendation to buy or sell any securities. Any companies mentioned do not\nnecessarily represent current or future holdings of any BlackRock products.
\n© 2026 BlackRock, Inc. or its affiliates. All Rights Reserved. BLACKROCK is a trademark of BlackRock, Inc. or its affiliates. All other trademarks\nare those of their respective owners.
\n 11