Pomegra
Recursive Superintelligence raised $650M at a $4.65B valuation, led by GV and Greycroft, to build AI that improves its own code and automates scientific discovery.
Key Takeaways
- Recursive raised about $650M at a $4.65B valuation, led by GV and Greycroft, with Nvidia and AMD's venture arm joining.
- The San Francisco startup was founded in 2025 and has fewer than 30 employees.
- Its first goal is an AI that rewrites its own code, then extends to physics, chemistry and pre-clinical biology.
Lead
Recursive Superintelligence, a San Francisco startup led by former Salesforce chief scientist Richard Socher, came out of stealth on May 13, 2026 with roughly $650M in funding at a $4.65B valuation. Alphabet's GV and Greycroft led the round, which is described as a Series A. Nvidia and the venture arm of AMD also participated. The company says it is building recursive self-improving superintelligence: a model that can propose experiments, run them, validate the results and feed what it learns back into its own development.
What Is Recursive Building?
Recursive is building an AI system that improves its own code base, and its first target is AI research itself. The model is meant to find weaknesses in its own software, in the auxiliary programs that surround it (the "harness"), and in its training and inference infrastructure. It then tests fixes through simulation in an open-ended loop.
Socher has said the company will begin with AI research and later widen its scope to physics, chemistry and, in particular, pre-clinical biology. He has compared the potential effect on biology to what calculus did for physics. He has also said products will arrive in "quarters, not years," which separates Recursive from pure research labs.
The method leans on open-endedness and "rainbow teaming," in which multiple AI systems repeatedly test and challenge one another across millions of iterations. The approach borrows from biological evolution, where agents adapt and counter-adapt.
Who Is Behind Recursive Superintelligence?
The founding group is drawn from the largest AI labs. Socher is CEO and previously founded the search company You.com. Co-founders include Tim Rocktäschel, a University College London professor and former principal scientist at Google DeepMind, and Yuandong Tian, a former director at Meta FAIR. Alexey Dosovitskiy, a co-author of the Vision Transformer, and Josh Tobin, an early OpenAI researcher who led Codex teams, are also co-founders. Jeff Clune, Caiming Xiong and Tim Shi complete the list. Peter Norvig, co-author of a standard AI textbook, is an adviser.
Headcount is below 30, split between San Francisco and London. That puts the implied valuation above $150M per employee, a figure that reflects the pedigree of the team and the cost of compute more than any revenue. Recursive has disclosed no product revenue.
How Does the Valuation Compare With Earlier Talks?
The final number came in above the figure circulating months earlier. Bloomberg reported in January 2026 that Recursive was in funding talks at a valuation of about $4B. The closed round at $4.65B suggests demand held through the four months it took to finalize, and that the lead investors were willing to pay roughly 16% more than the early talks implied.
The round is also large for a company that had not publicly shipped anything. A $650M check to a team of fewer than 30 people mostly buys compute and researcher retention. Neither is cheap, and both are contested by better-funded rivals.
Why Are GV, Nvidia and AMD Backing It?
The investor mix mirrors where the money in AI flows. GV brings Alphabet's research lineage, and Greycroft adds a lead from a conventional venture firm. Nvidia and AMD's venture arm are chip suppliers whose revenue depends on labs spending heavily on training compute. Their participation ties capital to future hardware demand, a pattern that has become common across large AI rounds and that complicates any reading of the valuation as pure market validation.
Can a Startup Compete With the Big Labs?
The competitive picture is difficult. Anthropic, OpenAI and Google DeepMind already pursue AI that accelerates its own research, and each has far more compute and capital than Recursive. A separate startup, Ineffable Intelligence, is also using reinforcement learning to build self-improving models.
The central technical question is whether recursive self-improvement produces runaway acceleration or runs into diminishing returns. Socher acknowledges that limits to intelligence exist but calls them "astronomical" and distant. Skeptics point out that gains in one part of the training stack, such as code, do not automatically translate into broader scientific discovery. Recursive has also said it plans to build in safety guardrails, but it has not published details of what they are.
What Comes Next for Recursive?
The near-term test is whether the company ships a working system that improves measurably on its own output, rather than a research demonstration. Socher's own "quarters, not years" timeline sets a public benchmark. If the first release shows compounding gains in AI research tasks, a follow-on round at a much higher valuation is plausible. If gains flatten, the company faces a harder case against rivals that can spend ten times as much on compute.
Expansion into biology is the longer bet. Pre-clinical research depends on wet-lab validation that a software-only system cannot supply, so partnerships with drug developers or lab automation firms would be a likely next step.
Outlook
Recursive combines a strong founding team, a $4.65B valuation and a thesis that the largest AI labs also pursue. The $650M gives it room to train at scale, but with fewer than 30 people and no disclosed revenue, its value rests on whether self-improvement loops deliver measurable progress within the timeline Socher has set. The next signals are a first product release and any evidence of gains that carry beyond AI research into the sciences.
