The First Move. The people and capital behind the AI race. |
AI's Best Researchers Stopped Taking Offers. Now They Take Term Sheets. Last issue, Big Tech bought AI teams. This week, the teams walked out, priced themselves in the billions and raised before shipping anything. Here's what that does to everyone else trying to hire them. Written by Jack Leyden | Issue 02 |
| Together with Accelera Growth Partners, helping AI founders learn, build, and scale. |
On August 5, Jeff Dean resigned as Google's chief scientist after 27 years building the infrastructure behind much of Google's search, ads and AI. He didn't leave alone. Sanjay Ghemawat, Quoc Le and Oriol Vinyals left with him to co-found Discovery Loop, a public benefit corporation built to automate the scientific method.
The money moved faster than the product. Weeks after the company was reportedly seeking $1 billion at roughly a $10 billion valuation, Business Insider reported it was in talks at around $50 billion. It has not announced a product, a customer or a launch date.
That is roughly a fivefold jump in weeks for a company with nothing to sell yet. The market isn't pricing a product. It's pricing four people.
The neolab playbook
The industry has a name for these companies now. A neolab is a research-first startup founded by people who left a frontier lab, a Big Tech AI group or a university to chase a long-horizon breakthrough, usually at a billion-dollar valuation before anyone outside can try anything.
Discovery Loop is the latest, not the first. Yann LeCun's AMI Labs raised $1.03 billion in March. David Silver raised a record $1.1 billion seed for Ineffable Intelligence. humans&, founded by ex-OpenAI researchers, raised a $480 million seed at a $4.5 billion valuation in January.
The price of a research team Latest disclosed round for each lab, and whether outsiders can use anything yet. Bar length is to scale.
Ten leading neolabs: about $11.6B in latest disclosed rounds. Two have shipped products. Source: VC Cafe neolab scorecard, September 23, 2026. Discovery Loop's reported $50B talks are not included. |
Thinking Machines Lab and World Labs have crossed the line into products outsiders can use. The rest are still asking investors to underwrite a team and a thesis. By VC Cafe's count, most neolabs have secured the team and the compute, a few have put something in outsiders' hands, and none has yet proven economics that justify the valuations.
In 2026, a research team is worth more before it ships than most companies are after.
Why investors write these checks
The push. The big labs stopped being research labs. Radical Ventures' head of talent argues ChatGPT changed the DNA of places like OpenAI: they are now product-driven companies fighting for market share, and neolabs are the response. Inside the incumbents, benchmark pressure and fast release cycles leave little room for exploratory work.
The pull. Investors now treat elite researchers as an asset class. The logic is downside protection: even if a neolab never commercializes, the team it assembled still has value and can land softly inside Big Tech.
That connects straight to Issue 01. The soft landing investors are counting on is the license-and-hire deal we broke down last week. The acquihire has become the floor that makes the neolab bet work. Big Tech sells the talent its exit, then buys the talent back.
Big Tech is adapting too. Rather than fight Dean, Alphabet is a founding investor and cloud partner in Discovery Loop and will supply its compute for the first year. If you can't keep your best researchers, own a piece of what they build next.
The revolving door spins both ways
Neolabs don't keep everyone either. Thinking Machines launched in February 2025 with six co-founders. Four have left: Andrew Tulloch to Meta, Barret Zoph and Luke Metz back to OpenAI in January, and Lilian Weng back to OpenAI in July to lead research on recursive self-improvement. Only Mira Murati and chief scientist John Schulman remain, with Soumith Chintala now CTO.
The lesson for founders: capital buys you a team, not loyalty. The same researchers who give a neolab its valuation can take it with them.
The squeeze in the middle
Researchers used to choose between a big lab and a startup. Now there is a third option, and it raises the bar for every offer. Frontier labs still win on pay and compute. Neolabs win on research freedom. Traditional AI startups get squeezed between the two.
If you run an applied AI company, you're competing for senior people against Big Tech cash, a billion-dollar research budget with no product deadline, and the chance for that researcher to start their own lab next year. You won't win that auction on money.
Five moves for founders hiring into this market
Sell the problem, not the paycheck. Neolabs win on freedom. Your counter is an important problem with real users and a shorter path to seeing the work matter. Lead with it in every senior conversation.
Hire the second circle. Neolabs recruit from the tiny pool that has trained frontier systems. The engineers who built evals, data pipelines and inference around those people are more gettable and often more useful to an applied company.
Offer ownership that feels like founding. Your real competitor is the founder route. Close the gap with meaningful equity, real scope and a title that means something.
Plan for your best researcher to leave and start a lab. Structure vesting and IP terms that are fair both ways, and consider being their first design partner or investor, the way Alphabet was with Dean.
Treat neolabs as partners, not just rivals. Most have capital and research but no product or customers yet. If you have distribution, you have something they need.
In AI right now, the team is the product. Big Tech pays billions to buy it and investors pay billions to back it. Price your people like the asset the market already says they are.
If you were a top researcher at a big lab today, what would you do?
Get The First Move every week
Free, one issue a week on the strategy behind AI talent, capital and deals.
Know a friend who would want this? Forward it to them, or share thefirstmove.beehiiv.com/subscribe.
Who's hiring: the neolabs
The labs in this issue are building teams right now. Most don't post pay, so treat the posted ranges as the floor.
San Francisco Bay Area
Thinking Machines Lab San Francisco. Sponsors visas | $350K–$475K |
Periodic Labs Menlo Park or Montreal. Sponsors visas | $250K–$350K plus equity |
Periodic Labs Menlo Park, with travel to Taiwan | $200K–$275K plus equity |
humans& San Francisco | Not posted |
Europe
Ineffable Intelligence London office | Not posted |
Pay as posted on each company's own job board, checked September 24, 2026. Equity and bonus not included unless noted.
25 researchers to watch in the Bay
The people whose next move, paper or departure will tell you where AI talent and capital go next.
Neolab founders and leaders
| 01 | Jeff Dean Discovery Loop, co-founder Left Google after 27 years to automate the scientific method. The biggest name yet to leave for his own lab. |
| 02 | Quoc Le Discovery Loop, co-founder Founding member of Google Brain, now betting that AI can run experiments at a scale no human lab can. |
| 03 | Mira Murati Thinking Machines Lab, CEO Runs one of the two neolabs that has actually shipped a model and a product. |
| 04 | John Schulman Thinking Machines Lab, chief scientist OpenAI co-founder whose reinforcement learning work sits underneath ChatGPT. |
| 05 | Soumith Chintala Thinking Machines Lab, CTO Co-creator of PyTorch, stepped in as CTO after Barret Zoph's exit. |
| 06 | Ilya Sutskever Safe Superintelligence, co-founder Leads the most highly valued lab with nothing public to evaluate. Every release, if one comes, will matter. |
| 07 | Fei-Fei Li World Labs, co-founder Building spatial intelligence. World Labs is one of the two neolabs with a product outsiders can use. |
| 08 | Justin Johnson World Labs, co-founder Research lead behind World Labs' world models and developer API. |
| 09 | Liam Fedus Periodic Labs, co-founder Former OpenAI researcher pairing AI scientists with autonomous physical labs. |
| 10 | Ekin Dogus Cubuk Periodic Labs, co-founder Materials science and machine learning researcher behind Periodic's lab-in-the-loop approach. |
| 11 | Karol Hausman Physical Intelligence, CEO Building general-purpose robot foundation models out of San Francisco. |
| 12 | Andrej Karpathy Eureka Labs, founder Former OpenAI and Tesla AI lead. What he explains publicly tends to set how the field talks about new ideas. |
Frontier lab researchers
| 13 | Lilian Weng OpenAI Back at OpenAI after co-founding Thinking Machines, leading research on recursive self-improvement. |
| 14 | Barret Zoph OpenAI Former Thinking Machines CTO, returned to OpenAI in January. A bellwether for the revolving door. |
| 15 | Noam Brown OpenAI Reasoning research lead whose work shaped OpenAI's reasoning models. |
| 16 | Mark Chen OpenAI, research leadership Senior research leader overseeing much of OpenAI's model work. |
| 17 | Chris Olah Anthropic, co-founder Pioneer of interpretability, the work of understanding what happens inside models. |
| 18 | Jan Leike Anthropic Leads alignment research after leaving OpenAI's superalignment team. |
| 19 | Jim Fan NVIDIA Leads embodied AI and robotics research at NVIDIA. |
Academia and robotics
| 20 | Percy Liang Stanford Leads Stanford's Center for Research on Foundation Models and its model evaluation work. |
| 21 | Chelsea Finn Stanford and Physical Intelligence Robot learning researcher splitting time between Stanford and a leading robotics startup. |
| 22 | Sergey Levine UC Berkeley and Physical Intelligence One of the most cited researchers in robot learning and reinforcement learning. |
| 23 | Pieter Abbeel UC Berkeley Robotics and reinforcement learning researcher and prolific investor in AI founders. |
| 24 | Christopher Manning Stanford Natural language processing pioneer who trained a generation of today's lab researchers. |
| 25 | Tatsunori Hashimoto Stanford Researches how language models learn and fail, and how to measure it. |
Our picks, weighted toward people building or leading research in the San Francisco Bay Area. Roles as last publicly reported, September 2026. People move fast in this market, which is the point.
Nothing here is legal or investment advice.
