Recursive Superintelligence Raises $650M at $4.65B — GV, Nvidia, and AMD Back a Lab With 25 Staff and No Product
Recursive Superintelligence emerged from stealth on May 13 with $650 million in funding at a $4.65 billion valuation. The company has roughly 25 employees, offices in London and San Francisco, and has not shipped anything to the public.
GV and Greycroft led the round. Nvidia and AMD Ventures participated. For context: this is more capital than Anthropic raised in its first three years, deployed into a company that has been operating for less than six months.
The Team Is the Product Pitch
The founding team is the reason investors moved at this speed. Richard Socher — former chief scientist at Salesforce, founder of You.com — is CEO. Tim Rocktäschel, formerly a principal scientist at Google DeepMind and professor of AI at University College London, co-founded alongside him. Jeff Clune, Josh Tobin, and Tim Shi add OpenAI and Uber AI research depth. Clune in particular is closely associated with open-endedness — systems that generate genuinely new capabilities over time.
The Financial Times had reported a figure north of $500M in April. The disclosed round at $650M with $4.65B valuation suggests oversubscription.
What Recursive Is Building
The thesis: the bottleneck in frontier AI is not compute or architecture but the research loop itself — the cycle of evaluation, data selection, training, post-training, and research direction. Recursive is trying to automate that loop. If it works, the constraint on building better AI shifts from hiring great researchers one by one to owning the process that generates better AI with less direct human steering.
The company is drawing on Stanisław Lem’s concept of an “information barrier” — the point where available knowledge grows faster than humans can meaningfully integrate it. Tim Rocktäschel frames Recursive’s goal as breaking through that barrier by fully automating the scientific method, starting with AI research.
Concretely: the system will develop experiment ideas, test them, validate results, and improve not just its own weights but its harness — the auxiliary programs that shape how a model is trained, evaluated, and deployed. Training and inference infrastructure are also in scope.
No Results Yet
Recursive has published no technical results. The company plans a wider launch in mid-2026 and expects to scale compute across its San Francisco and London sites. Guardrails for preventing risky output are described but not specified.
The round tells something specific about where venture capital is in mid-2026: a team of recognizable frontier AI researchers with a credible thesis commands institutional-scale funding before a single benchmark or user. Startups with less elite pedigree — but working products — raise at smaller valuations and face more scrutiny.
The immediate question is whether Recursive can produce something measurable before the next wave of self-improvement papers arrives from the labs it drew talent from.