A UK-based AI startup with fewer than 30 employees, no published research, and no shipping product just raised $650 million at a $4.65 billion valuation. Its name is Recursive Superintelligence. And yes, it means exactly what it sounds like.

The company emerged from stealth on Tuesday with a thesis that would have sounded like science fiction two years ago: build AI that recursively improves itself, starting with AI research, then expanding to all of science. Co-founders Richard Socher, formerly Salesforce AI chief, and Tim Rocktaschel, formerly of Google DeepMind, have assembled a team of researchers from OpenAI, Meta AI, and Uber AI to chase what they call "the fastest path to superintelligence."

The $650 million round was led by GV (formerly Google Ventures) and Greycroft, with AMD Ventures and Nvidia also participating. In April, the Financial Times reported the round at $500 million. The final number came in 30% higher.

The Thesis: AI That Automates Science

Rocktaschel draws on Stanislaw Lem, the Polish science fiction writer, and his concept of an "information barrier," the point where available knowledge grows so fast that humans can no longer meaningfully integrate it. The company believes AI has already crossed this barrier in machine learning research. Papers publish faster than any researcher can read them. Benchmark results obsolete within weeks. The field is too fast for human cognition alone.

Their plan: fully automate the scientific method, starting with AI research itself. Build an AI that can read every paper, run every experiment, identify every dead end, and propose every next step. Then let it improve its own architecture. Then point it at biology, chemistry, physics.

"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 announcement.

The Numbers Do Not Add Up. That Is the Point.

A $4.65 billion valuation for a company with no product, no revenue, no published research, and fewer than 30 employees would have been absurd in any previous era of technology investing. In the current AI market, it is not even the most aggressive bet this month. Anduril raised $5 billion at $61 billion on Tuesday. Anthropic is in talks for $50 billion at a $950 billion valuation.

What makes Recursive different is the explicitness of the ambition. Most AI companies talk about building useful tools. Recursive talks about building a system that replaces the need for human researchers. The name itself is a declaration of intent: recursive self-improvement is the theoretical mechanism by which an AI system could rapidly surpass human intelligence once it reaches a threshold of capability.

The investor list is telling. Nvidia and AMD are not betting on a chatbot competitor. They are betting on a company that will need enormous amounts of compute to run recursive improvement loops. GV, Google's venture arm, is backing a company whose success could render Google DeepMind's entire research organization redundant.

The Safety Question Nobody Is Answering

Recursive self-improvement is the scenario that keeps AI safety researchers awake. It is the mechanism behind most theoretical models of uncontrollable AI. When Anthropic released Mythos earlier this year, triggering White House emergency consultations and national security reviews, the concern was that frontier models were becoming powerful enough to pose novel risks. Recursive Superintelligence is building a company around making that concern its product.

The company has not published any safety research. It has not described alignment protocols. It has not explained how a self-improving system would remain controllable. In its announcement, the word "safety" does not appear.

This is happening while the Musk v. OpenAI trial in Oakland enters closing arguments on Thursday over whether AI companies have a responsibility to prioritize safety over profit. While Illinois pushes AI regulation bills through the state legislature. While the Trump administration scrambles to decide whether to regulate AI at all.

What to Watch

Recursive plans a public launch in mid-2026. That is weeks away. When it ships, the AI industry will get its clearest test yet of whether recursive self-improvement is a real engineering approach or a fundraising narrative. The company's founders have genuine pedigrees. But so did the founders of every well-funded AI company that has struggled to turn research ambition into product reality.

The question is not whether $650 million can buy good researchers. It can. The question is whether the thing they are building can be controlled once it works.