Richard Socher sat for an interview on April 9. His company had existed for 99 days. About 20 employees. No product. No public website worth visiting. No customers.

Twelve days later, he closed a $500 million round at a $4 billion pre-money valuation. The round was oversubscribed to nearly $1 billion. GV led. Nvidia wrote a strategic check.

The Pitch

Recursive Superintelligence was incorporated on December 31, 2025. Four months later, it's worth more than most publicly traded biotech companies with FDA-approved drugs sitting in pharmacies.

The pitch is simple: build an AI that builds AI. Automate the frontier development pipeline — evaluation, data selection, training, post-training, even the choice of research direction. Feed the system back into itself. Let it improve. Step back. Watch the curve bend.

The founding team is real. Richard Socher was chief scientist at Salesforce. Tim Rocktäschel led a team at Google DeepMind that won ICML 2024's Best Paper Award for the Genie interactive world model. Josh Tobin, Jeff Clune, and Tim Shi came out of OpenAI. These are not first-time founders with a LinkedIn post.

The target market is also real. Frontier labs burn an ugly share of researcher time on work a well-designed system could automate — hyperparameter sweeps, eval harnesses, data curation, failed training runs nobody remembers. Automate the researchers themselves and you compound. That's the thesis. That's what GV and Nvidia just paid $500 million to find out.

Why This Is Strange

Notice what's missing from that paragraph. Progress. Benchmarks. Demos. Anything resembling a product. The company plans its public launch around mid-May. GV and Nvidia didn't wait for the launch. They funded the idea and the résumés, full stop.

This is what the late stage of a capital cycle looks like. In 2021, it was crypto. In 2024, it was any LLM wrapper with a tasteful landing page. In 2026, it's founder pedigree plus the word "recursive" somewhere in the deck. Diligence compressed from months to weeks to — apparently — whatever Socher said over lunch.

My Opinion

I'll be blunt: I don't mind the bet. I mind the price.

Self-improving AI is a real research direction with enormous upside if it works. Socher's team is credentialed. Nvidia's participation signals they want preferred access to whatever comes out — chips for code, probably. Fine. That's venture capital doing its job.

But $4 billion for 20 people and a slide deck is something else. It's a statement about how much capital is chasing how few AI teams who can credibly pitch a Big Idea. Anthropic just paid $400 million for nine researchers earlier this month. Meta dropped $14 billion on a Llama replacement that didn't land. Now a four-month-old lab with no product is worth more than Quora. The pattern isn't "great teams get funded." The pattern is "scarcity of credentials meets a wall of money." Prices decouple from fundamentals because there are no fundamentals yet to decouple from.

What I actually want to see is the term sheet. Liquidation preferences. Ratchets. Board control. Because "oversubscribed at $4B" is the part investors want you to read. The protections they negotiated are the part that decides whether this ends in a useful research lab or a 2028 retrospective about how smart people convinced themselves a pitch deck was a company. In May they'll launch something. I'll be watching. So will GV's LPs, who just became reluctant partners in the thesis that an AI can train itself to be smarter than the team that made it.

Author: Yahor Kamarou (Mark) / www.humai.blog / 21 Apr 2026