Companies·2 min read·SiliconANGLE

Richard Socher’s Recursive Superintelligence Exits Stealth With $650M to Build AI That Improves Itself

Backed by GV, Greycroft, Nvidia, and AMD at a $4.65B valuation, the Socher–Rocktäschel lab is betting recursive self-improvement is the fastest path to superintelligence.

Richard Socher’s Recursive Superintelligence Exits Stealth With $650M to Build AI That Improves Itself
Share:

Recursive Superintelligence emerged from stealth this week with a $650 million funding round at a $4.65 billion valuation, instantly placing it among the best-capitalized AI research labs in the world. The San Francisco and London-based startup is co-founded by former Salesforce chief scientist Richard Socher and former Google DeepMind researcher Tim Rocktäschel, with a team of more than 25 researchers and engineers drawn from OpenAI, DeepMind, Meta, Salesforce, and Uber AI.

The round was led by GV and Greycroft, with participation from Nvidia and AMD Ventures—an unusual pairing that signals both chip giants are willing to seed customers chasing the bleeding edge. “The fastest path to superintelligence will be realized by AI that recursively improves itself, and does so via open-ended algorithms that drive endless innovation,” the company said in its launch statement. Rocktäschel has invoked Stanisław Lem’s notion of an “information barrier,” arguing that human researchers can no longer keep up with the volume of scientific output and that automating the scientific method itself is the only way through.

Recursive’s plan is to start by automating AI research—training systems that can analyze their own performance, generate new architectures and training recipes, and run experiments without human direction—then extend that loop into other scientific domains. That puts it squarely in the same conceptual neighborhood as OpenAI’s superalignment work, DeepMind’s recent co-mathematician results, and Anthropic’s automated red-teaming, but with a corporate structure built around self-improvement as the explicit product rather than a side effect.

The lab has not yet published any concrete technical results, and a public launch is planned for mid-2026 as it scales compute infrastructure across its two offices. Even so, the optics of Nvidia and AMD writing checks into the same round—alongside a leadership team with a track record of shipping at Salesforce and DeepMind—will make Recursive one of the most closely watched debutants of the year. If self-improving AI works as advertised, the team building it stops needing to hire more humans; if it doesn’t, this becomes the most expensive bet yet that the bitter lesson has another chapter.

Comments

Share your thoughts. Be kind.

0/2000

Loading comments…

Related Articles

SPACEX · NASDAQ: SPCXLARGEST IPO IN HISTORY$1.75TTarget valuation · up to $75B raisedSurpassing Saudi Aramco's $35.4B record raiseJUN 4RoadshowJUN 11PricingJUN 12Trading · SPCXSource: SpaceX S-1 · CNBC
Companies

SpaceX Sets Terms for the Largest IPO Ever: ~$1.75T Valuation, Up to $75B, Trading June 12 as SPCX

AI SILICON · THE CHIP RACEMICROSOFT × ANTHROPIC4thSilicon vendor in Anthropic's stackif the Maia 200 talks with Microsoft closeStatus: early talks — nothing signed yetMICROSOFT SILICONMAIAMAIA 200216 GB HBM3e · TSMC 3 nm10+ PFLOPS FP4 · launched Jan 2026WHAT IT MEANSAlready on AWS, Google and Nvidia chipsMicrosoft holds a ~$5B stake in AnthropicSource: WSJ / CNBC
Companies

Anthropic in Talks to Run Claude on Microsoft's Maia 200, Adding a Fourth Silicon Vendor

TeamPCP Hackers Put 450 Mistral AI Repos and 5GB of Source Code Up for Sale at $25,000
Companies

TeamPCP Hackers Put 450 Mistral AI Repos and 5GB of Source Code Up for Sale at $25,000