SAN FRANCISCO: Recursive Superintelligence exited stealth on May 13, 2026 with $650 million in funding at a $4.65 billion valuation, one of the largest stealth-to-public rounds of the current AI cycle. The round was led by GV and Greycroft, with participation from AMD Ventures and Nvidia, according to TechCrunch’s reporting. The company is led by Richard Socher, the former chief scientist at Salesforce and founder of AI search engine You.com.
The team formed roughly four months before the announcement. The thesis is sharper than the timeline: build AI systems that autonomously identify their own weaknesses, redesign their own architecture and training methods, and improve themselves without continuous human oversight. The company is calling the approach recursive self-improvement.
Who founded Recursive Superintelligence?
Socher is CEO. The seven co-founders are Tim Rocktäschel, Alexey Dosovitskiy, Josh Tobin, Caiming Xiong, Yuandong Tian, Tim Shi, and Jeff Clune. The collective resume reads like a frontier-AI all-star draft. Rocktäschel is a professor of AI at University College London and a former DeepMind research director who led open-ended learning work. Dosovitskiy is a co-author of the Vision Transformer paper. Yuandong Tian was a research director at Meta’s Fundamental AI Research lab. Clune, Tobin, and Shi come out of OpenAI. Xiong worked under Socher at Salesforce. Peter Norvig, co-author of the standard AI textbook Artificial Intelligence: A Modern Approach and a longtime Google research director, signed on as an adviser, as detailed by Unite.AI’s writeup of the round.
Eight co-founders is unusual at any stage. Eight co-founders this senior, four months in, is a different category. It suggests Socher recruited people who already knew each other’s work, agreed on the technical bet, and were willing to ship equity across a wide table to lock the team before competitors could split it. For comparison, recent co-founder dynamics at xAI and other frontier labs have shown how fragile these teams get once the work starts.
How much did Recursive Superintelligence raise and at what valuation?
The headline figure is $650 million at a $4.65 billion post-money valuation. That puts the company in the top tier of 2026 AI rounds by absolute size, and at the very top when measured against employee count. According to SiliconAngle’s coverage, the company plans to use the capital to secure large-scale compute infrastructure and run its first “Level 1” autonomous training system, with a public launch targeted for mid-2026.
The cap table is the more interesting story. GV and Greycroft are tier-one growth investors with deep AI portfolios. AMD Ventures and Nvidia participating together on the same round is the structural detail to flag. The two chipmakers rarely co-invest in frontier AI. Both writing checks here signals neither wants to be on the wrong side of a recursive self-improvement breakout, and both are willing to subsidize a startup that will buy capacity from whichever silicon performs. For founders watching the AI infrastructure layer, that is a marker of how seriously the compute supply chain is treating the research-automation thesis.
What does Recursive Superintelligence actually do?
The technical pitch is research automation. Today, frontier AI progress is bottlenecked by humans: researchers design architectures, write training code, design evaluations, and decide what to try next. Recursive’s bet is that an AI system can do all of that work on itself, faster than any human team, and that the compounding gains from each iteration will eventually outpace human-led research. The Decoder’s technical overview describes the system independently altering its own neural network architecture, refining its training algorithms, and managing evaluation protocols without continuous human oversight.
It is a thesis that has lived mostly in academic papers and DeepMind whiteboards for two decades. Clune’s earlier work on the Darwin Gödel Machine at Sakana AI is one of the cleaner precedents. Recursive’s claim is that the technique is now tractable at production scale, given enough compute and the right team. Socher told press the company plans to ship products in “quarters, not years.”
What this means for founders
The investor narrative around this round will fixate on the existential-risk framing. The more useful read for founders is structural: the next billion-dollar moat in AI may not be in foundation models. It may be in the layer above them, where research itself gets automated. If Recursive’s bet works, the moat shifts from “we trained a bigger model” to “we built the system that decides what to train next.” Foundation-model labs become customers of that layer, not its peers.
A few practical takeaways. First, the speed signal. Four months from formation to $650M is the fastest stealth-to-megaround timeline in recent memory, and it tells you the market is paying for team caliber and thesis, not traction. If you are recruiting AI research talent right now, your competition just got steeper and better capitalized. Second, the compute angle. Recursive will be buying compute at the rate of a frontier lab without ever shipping a consumer product, which tightens supply for everyone else. Expect ripple effects in capacity availability through the rest of the year, on top of the constraints already visible in Anthropic’s recent SpaceX Colossus deal and the broader Google Cloud backlog. Third, the cap table lesson. When Nvidia and AMD both back the same startup, the smart-money read is that both think the bet is large enough that picking sides early is more expensive than buying optionality.
The closer milestone to watch is the mid-2026 product launch. If Recursive ships something concrete on Socher’s “quarters, not years” timeline, the research-automation layer becomes a real category. If the launch slips or the first system underperforms, the thesis goes back into the academic-paper pile and the cap table absorbs the loss. Either way, the round itself has already moved the AI conversation up one level of the stack.
