Recursive Superintelligence bets 63% of its cash on a $410M AWS compute deal

Recursive Superintelligence just made the kind of bet that separates ambitious AI labs from the merely well-funded ones. The company, which burst out of stealth in May with $650 million, announced Tuesday it will spend $410 million on cloud compute through Amazon Web Services. That’s 63% of its total war chest, plowed directly into raw processing power.

Founder Richard Socher doesn’t sound worried about the burn rate. On a call with TechCrunch, he called this agreement “likely going to be one of the smallest compute deals we’re going to sign in the next few years.” The math is brutal and deliberate. Recursive is chasing recursive self-improvement — AI systems that get better without human engineers hovering over them — and that approach substitutes expensive GPUs for expensive salaries. Socher put it bluntly: “For us, it’s less about headcount and more about agent count.”

Unlike the convoluted investment-and-cloud-credit packages that have become standard for frontier labs, this deal contains no equity component. Amazon isn’t taking a stake. What AWS gets instead is a flagship customer with extreme infrastructure demands — the kind of stress test that forces a cloud provider to build better plumbing. Jason Bennett, AWS’s VP for startups, said the two companies will “co-develop infrastructure purpose-built” for AI firms operating at this scale. It’s a reference account that might convince other foundation-model companies to sign similar deals.

Socher says the first products from this compute-heavy approach will arrive sooner than skeptics expect. He pointed to October as the moment when “tangible, useful things” will be available for people to try. Self-improving AI has been a theoretical holy grail for years, with timelines ranging from imminent to never. Recursive is now spending hundreds of millions to force an answer — and we’ll know within months whether that bet was prescient or premature.