Neolabs today are being asked to pull off two moonshots at once. Their investors want both a research breakthrough and a venture-scale business, even though each of those alone is a 1 in 100 outcome. Asking for both makes their chances 1 in 10,000.
I think there’s a better alternative, one that borrows from the way pharma handles risky R&D and lets frontier labs, neolabs, and investors all come out ahead.
Frontier labs want to scale up what works
Today, OpenAI, Anthropic, and Google DeepMind are caught in a neck-and-neck race on model quality and intelligence. Every release needs to exceed expectations, or at least meet what competitors are doing. So they’d rather spend their resources scaling things that are proven to work, like improving data pipelines, building infrastructure for larger runs, and buying ever-more compute. However, researchers often want to work on “the next big thing”, new paradigms that could lead to a discontinuous jump in intelligence. As a result, there’s often tension between what the market needs them to ship and what researchers want to focus on.
Neolabs are asked to do two things at once
In the past year, many senior researchers who want to pursue these bets have started neolabs, nascent AI startups focused on big research bets, usually funded with a lot of money before they have a product or revenue, so they can pay for compute and salaries. That way they can chase the big idea with venture-scale resources.
On the surface, investors are betting that a neolab becomes the next OpenAI or Anthropic. They remind themselves that Anthropic was once a neolab too, started by a group of researchers who left OpenAI and raised hundreds of millions of dollars before they had a product. Privately though, most of them will admit the odds of a repeat are next to none, even with world-class researchers and engineers.
Running a research team is a different job from scaling a company. Researchers usually aren’t great at product, go-to-market, operations, logistics, etc., all the things that traditionally matter for a sustainable company. (Ali Ghodsi at Databricks is a good exception.)
However, neolabs are expected to do both: make a breakthrough and scale a product. Sometimes this splits a company in two, with half the people wanting to build and sell and the other half wanting to do research and build ASI. Focus and alignment are the biggest advantages a startup has over better-funded incumbents, but this expectation can take a toll on both.
Investors who expect their neolab to do both will likely be disappointed when it does neither.
One radical move is to admit that a neolab is just going to do world-class, groundbreaking research. Expect close to zero revenue and let them cook. If they succeed, they get acquired or acquihired.
The catch is that this doesn’t work for investors. A VC needs companies that can return the entire fund, and it’s hard to promise LPs that an acquihire will do that, especially at the sky-high multibillion valuations neolabs raise at today.
Pharma already solved a version of this
Drug development has dealt with a similar problem for decades. Companies regularly risk hundreds of millions of dollars in R&D on a drug that might pay off billions. A drug can pass early trials and still fail before FDA approval, usually because it doesn’t work well enough, and sometimes because of side effects or manufacturing problems.
Pharma handles this by splitting the work. Smaller biotech companies take on the risk of discovery and development. If a drug works, a big pharma company acquires the startup or buys the drug as IP. The big company doubles down on what it’s good at, which is producing and distributing drugs at scale. The small company gets to do the science without having to invent a business model. A small biotech’s business model comes down to one question: can we make this drug work or not?
The division of labor works. Of the drugs the FDA approved from 2013 to 2022, small companies with under \$500M in revenue originated 52%, while the biggest companies, with over \$10B in revenue, originated 36%<sup>1</sup>.
Biotech investors have made money this way for decades. Many of their biggest wins come from getting acquired, not from growing into the next Pfizer.
Pre-registered acquisitions
What’s missing today is a promise. The people who start, join, or fund a neolab need to know, before they take the risk, that a big payoff is waiting if the research works.
Frontier labs could make that promise ahead of time through pre-registered acquisitions. This kind of promise is also called “pull funding”. Push funding pays for research up front, like a grant or a VC round. Pull funding promises to pay for the result once it exists. In 1714, the British government promised £20,000 to anyone who could find a ship’s longitude at sea, and a self-taught clockmaker named John Harrison responded by inventing a clock that kept accurate time on a rolling ship. In 2009, five countries and the Gates Foundation promised \$1.5B to any company that could supply poor countries with pneumococcal vaccines at \$3.50 a dose or less, and vaccine makers responded with enough supply to immunize more than 150 million children.
Big companies have always bought startups for their patents or their people. What’s different about AI research is the cost of finding out whether an idea works. A software founder can test an idea on a laptop. Testing a new training method at a scale that matters takes a training run that can cost tens or hundreds of millions of dollars, and the cost of the largest runs has grown 2.4x a year since 2016. Nobody spends that much on a guess without knowing what success is worth. The stakes are also bigger than any one company’s exit. A breakthrough in how models learn ends up in systems that hundreds of millions of people use every day, so how we pay for that research decides how fast it happens and who gets to do it.
I can see this playing out in two ways:
- Open offers. A frontier lab publishes a target anyone can check and commits to acquiring the first team that hits it, at a set price. OpenAI’s Parameter Golf challenge is a small version of this, with job interviews for standout entrants instead of an acquisition.
- Private options. A frontier lab and one neolab agree on a scoped goal up front. The lab gets the right to buy the neolab at a set price if the goal is met, and pays a fallback fee if it walks away. SpaceX’s option to buy Cursor had this structure: \$60B if SpaceX bought, or \$10B for the work if it didn’t. SpaceXbought it .
In both cases, the target can be a new capability, like a benchmark score, or something the lab can’t easily build itself, like a dataset, a set of RL environments, or a regulatory approval.
Here’s what an open offer could look like. Say a frontier lab posts: “We’ll acquire any team that matches our last-generation model on this eval suite using a tenth of the training compute, for \$2B. Train from scratch, no distilling from anyone’s frontier model, and the offer is good through 2027.” The price makes sense, because the largest training runs are on track to cost more than \$1B each by 2027, so cutting the compute for every future run by 10x is worth far more than \$2B.
Why would a frontier lab commit?
The obvious question is why a lab would name a price before seeing results, when it could wait and buy whatever works.
After a breakthrough, it’s too late. Training recipes can’t be patented, so once a team shows a lab how it did something, the lab doesn’t need the team anymore. No sensible team will show its work without a price agreed first, and no lab will pay for something it hasn’t seen. This is called Arrow’s information paradox. Naming the price in advance breaks the deadlock: the team knows what it gets, and the lab gets to check the result before it pays.
A win-win-win solution
Pre-registered acquisitions let everyone come out ahead.
The frontier labs get to:
- Keep scaling what works.
- Bet on 0-to-1 breakthroughs without funding every long shot themselves.
Neolabs and their researchers get to:
- Chase unproven but high-potential bets that could lead to a breakthrough, with funding behind them.
- Skip product-market fit and go-to-market, and spend their time on the research they’re best at and most excited about.
- Stay internally aligned, because everyone knows they can win just by making a breakthrough.
- Form a team, or go solo, and take a shot at a published target. Smaller targets work too: if a lab posts \$100M for a new set of RL environments, spending \$100k on compute to try is a reasonable bet.
Investors get to:
- Take on less market risk, because if their team gets there first, there’s a committed buyer at a known price.
- Value a neolab against the acquisition price. That won’t justify a \$5B seed round, but it will justify a \$100M one, and a \$2B exit on a \$100M entry returns the fund.
An open offer is also a floor, not a ceiling. If a team hits the target and would rather build a business around what it made, it can turn the offer down.
Making the implicit explicit
Some version of this already exists. Every past acquisition and acquihire, like Google’s \$2.4B Windsurf deal, sets an implicit anchor for what a research team can exit for, and neolabs point to those anchors to justify their valuations. A pre-registered acquisition just makes that explicit, reducing the uncertainty for everyone.
Right now a neolab has to win at research and at business, two moonshots at once. Put a price on the breakthrough and it only has to win one. The payoff is smaller than becoming the next Anthropic, but the odds go from 1 in 10,000 to 1 in 100, and that’s a bet you can build a team around.
So, frontier labs: what would you pay for a training run that costs a tenth as much? Name a price.
- One caveat: A drug is easy to patent, and an AI training recipe mostly isn’t . So in AI the payoff can’tjust be a sale of IP. It has to be an acquisition, either of the team itself or of a moat that’s hard to copy, like regulatory approvals, unique data, or unique physical infrastructure.↩︎