Recursive Superintelligence raises $650 million to pursue self‑improving AI

Recursive Superintelligence emerged from stealth on May 13 with a headline‑making $650 million Series A round that places its valuation at $4.65 billion. The money came from a mix of venture capital and chipmakers: GV, Alphabet’s investment arm, led the round alongside Greycroft, with strategic participation from Nvidia and AMD. The investors’ involvement signals a belief that the company’s core thesis—recursive self‑improvement—could become a near‑term compute customer for hardware suppliers.

The startup’s leadership reads like a résumé of the AI elite. Richard Socher, former chief scientist at Salesforce and founder of You.com, serves as CEO. He is joined by seven co‑founders, including Yuandong Tian, a former director of research at Meta’s FAIR lab; Tim Rocktaschel, a former principal scientist at DeepMind; Alexey Dosovitskiy, co‑author of the Vision Transformer paper; and Josh Tobin, who previously worked at OpenAI. Peter Norvig, co‑author of the standard textbook "Artificial Intelligence: A Modern Approach," sits on the advisory board.

Recursive Superintelligence’s mission is to create AI systems that can discover knowledge, continuously optimize themselves and evolve without human intervention—essentially mirroring biological evolution but on a timescale of weeks instead of millions of years. The company describes its roadmap in stages. The first phase will train a model with the combined expertise of roughly 50,000 doctors, allowing the system to conduct autonomous scientific research. A public "Level 1" autonomous training platform is slated for a mid‑2026 launch, and the raised capital will fund the massive compute infrastructure required for those experiments.

While other leading labs already employ AI to accelerate their own research—Anthropic’s Claude writes code, OpenAI’s GPT‑5.5 has improved token‑generation speed, and DeepMind’s AlphaEvolve targets scientific discovery—none have organized an entire commercial venture around recursive self‑improvement. Recursive Superintelligence positions the self‑improvement loop itself as the product, betting that the first company to master it will gain an exponential lead over competitors.

The market’s appetite for the idea is evident. With fewer than 30 employees and no product in the hands of customers, the startup’s $4.65 billion valuation reflects investors’ willingness to pay a premium for the possibility of runaway AI acceleration. Industry insiders remain divided on whether recursive improvement will yield ever‑increasing returns or hit diminishing‑returns ceilings. Anthropic co‑founder Jack Clark estimates a 60 percent chance that a fully autonomous training system will exist by the end of 2028, and a 30 percent chance by 2027.

For now, Recursive Superintelligence’s biggest hurdle is proving that the theory can be turned into practice. If successful, the company could reshape the economics of AI development, turning the lengthy, human‑driven research cycle into a rapid, self‑sustaining loop. Until then, the $650 million bet stands as a bold statement of confidence in a vision that, until recently, lived only in academic folklore.