Recursive Superintelligence signs $410M AWS deal to run self-improving AI at scale

Recursive Superintelligence has signed a $410 million multi-year compute agreement with Amazon Web Services, the company announced Tuesday July 28. The deal gives the startup access to the infrastructure it says is necessary to run AI systems designed to improve themselves — autonomously modifying their own architecture and training algorithms without human engineers adjusting the parameters.

The company's CEO Richard Socher, formerly chief scientist at Salesforce and one of the most-cited researchers in natural language processing, co-founded Recursive with alumni from OpenAI, Google DeepMind, Meta AI, and Uber AI. The startup emerged from stealth in May with $650 million in funding and a valuation of $4.65 billion. Backers include GV (Google Ventures), Greycroft, AMD Ventures, and NVIDIA.

The Self-Improving AI Thesis

Recursive's central bet is that AI systems can eventually contribute to their own development — analyzing their performance and helping design improved versions of themselves, creating a recursively accelerating improvement cycle. It's a concept that has remained largely theoretical in the industry; Recursive claims to be building infrastructure to make it operational at scale.

The company says its automated AI research system has already outperformed a two-year human leaderboard record and achieved state-of-the-art results on three benchmarks: NanoChat, NanoGPT Speedrun, and SOL-ExecBench. The implication is that compute budget, rather than headcount, becomes the primary scaling lever — which explains why $410M of a $650M raise goes straight to cloud infrastructure rather than salaries.

Why AWS

Jason Bennett, VP and Global Head of AWS Startups, described the fit in terms of parallelism: "AWS gives them elasticity to run loops in parallel at massive scale, turning months of sequential research into days of parallel discovery." The co-development of tailored infrastructure — part of the deal terms — suggests Recursive is looking beyond off-the-shelf compute toward custom acceleration for its research loops.

Socher's own framing of the deal was deliberately ambitious. The $410 million, he said, is "likely going to be one of the smallest compute deals we're going to sign in the next few years" — positioning Recursive among the tier of AI labs for whom compute scale is not a constraint but a strategic weapon.

The Broader Context

The deal lands against a backdrop of unprecedented AI infrastructure investment. From hyperscaler capex buildouts to startup compute agreements, 2026 has seen more capital flow into raw compute than any prior year. What makes Recursive distinct is the specific claim that its compute budget is not powering conventional model training — it's powering an automated research process that may, if the thesis holds, generate more efficient models than traditional human-directed approaches.

Whether that bet pays off remains to be seen. But $410 million in committed AWS compute — with more deals anticipated — suggests Recursive's investors believe the self-improving AI timeline is closer than conventional wisdom suggests, as first reported by TechCrunch.