AWS just signed one of its more unusual cloud deals yet, and the startup on the other end thinks it is just getting started.
Specifically, Amazon announced a multi-year cloud agreement worth $410 million with Recursive Superintelligence, an AI startup focused on something called recursive self-improvement, essentially AI systems that refine their own performance through repeated experimentation, largely without a human steering each iteration. The deal isn’t a standard compute lease either. Instead, both companies say they’re jointly building infrastructure specifically designed to support the kind of research Recursive is running.
Recursive itself is barely two months out of stealth. The company surfaced in May carrying a $4.65 billion valuation, with researchers who left OpenAI, Google DeepMind, Meta AI, and Uber AI founding and backing it, and it has raised $650 million total from investors including AMD Ventures and Nvidia. In an interview with TechCrunch, CEO Richard Socher said this AWS deal will likely end up being the smallest computing contract the company signs over the next few years, adding that what matters most to Recursive is what he called agent count rather than headcount, meaning the company plans to pour resources into compute rather than traditional hiring.
AWS confirmed there’s no equity stake attached to this deal, a notable departure from how many AI labs structure their cloud partnerships these days. Jason Bennett, AWS’s global VP of startups and venture capital, said the collaboration includes co-developing infrastructure built around a specific challenge: self-improving AI systems demand more computing power with every research loop, which makes parallel processing and infrastructure flexibility genuinely competitive advantages rather than nice-to-haves.
Meanwhile, Recursive says its automated research system has already beaten human-held records that stood for two years across three benchmarks, including NanoChat and NanoGPT Speedrun. Socher framed the underlying logic simply: AI can write code, and if you get that feedback loop right, you can essentially automate the scientific method itself. Bennett echoed that sentiment, saying Recursive delivered in a few months what community-driven optimization efforts took years to achieve.
Still, whether that kind of benchmark performance generalizes into something broader remains an open question, and reproducibility, safety, and cost efficiency will likely need more scrutiny before anyone calls this a genuine inflection point. For now, Recursive plans to unveil something usable around October, aiming to move from research demonstrations to an actual product people can touch.
