There is a new entrant in the race to build superintelligence, and it is not being subtle about its intentions.

Recursive Superintelligence, a London-based lab incorporated on December 31, 2025, just closed at least $500 million in pre-Series A funding led by GV (Google's venture arm) with Nvidia joining as a strategic investor. The valuation: $4 billion. The round was reportedly oversubscribed and could approach $1 billion. The headcount: roughly 20 people. The product: none.

Let me say that again. Twenty people. No product. Half a billion dollars. Four billion dollar valuation. Founded four months ago.

The Team Is the Product

The founding lineup explains the check. Richard Socher was Salesforce's chief scientist. Tim Rocktaschel directed research at Google DeepMind and holds a professorship at UCL. Josh Tobin and Jeff Clune have OpenAI and DeepMind pedigrees. Tim Shi rounds out the team. Between them, they have touched nearly every major frontier AI system of the last decade.

Their pitch is simultaneously the simplest and most ambitious in AI: build systems that improve themselves. Not chatbots. Not enterprise agents. Not image generators. Systems that automate the entire AI development process: evaluation, data selection, training, post-training, and research direction. No human in the loop at each step.

In plain English: they want to build AI that makes better AI, which then makes even better AI. The recursive part of the name is not a metaphor.

The Math That Terrifies Competitors

Here is the thing that should keep every other AI lab's leadership awake tonight. If self-improving AI compounds even 1.5x per improvement cycle, your 12-month product roadmap is worthless. The model family you locked your enterprise contracts into could be obsolete before the ink dries. Every API provider's pricing, every capability forecast, every competitive moat built on model performance becomes a moving target that moves faster than humans can plan for.

The capital intensity is the other signal. A $500 million pre-Series A for a four-month-old company with zero revenue is not venture capital in any traditional sense. It is a bet that the team capable of cracking recursive self-improvement gets to define the next era of computing. GV and Nvidia are not funding a startup. They are buying a lottery ticket to the singularity.

The Competition Is Already Stacked

Recursive Superintelligence is not alone. Yann LeCun's AMI Labs raised $1 billion to build world models. Ineffable Intelligence, founded by David Silver (the DeepMind researcher behind AlphaGo), is chasing reinforcement learning breakthroughs from London. OpenAI, Anthropic, and DeepMind all have internal teams working on automated AI research. The difference is that Recursive Superintelligence has made it the entire company's reason to exist.

The company plans a public launch around mid-May 2026. Until then, we have a $4 billion valuation attached to a promise, five resumes, and a name that sounds like it was generated by an AI trying to scare regulators.

The Bigger Picture

This raise tells you everything about where AI capital is flowing right now. Q1 2026 saw $242 billion in AI venture funding. Four companies took 65% of all global VC. The capital is concentrating at the frontier, and the frontier just got another order of magnitude more expensive to compete at.

If Recursive Superintelligence works, it obsoletes every AI lab that relies on human researchers to make progress. If it does not work, it joins a growing list of billion-dollar AI bets that were really just very expensive talent acquisitions.

Either way, GV and Nvidia just told you what they think the ceiling is. And the ceiling does not have human researchers in it.

First reported by the Financial Times. Additional reporting from The Decoder, TechFundingNews, and Asanify.