Recursive signs $410M AWS deal to power self-improving AI research
On 28 July 2026 Recursive Superintelligence announced a multi-year, $410 million agreement with Amazon Web Services, which will become the cloud provider running its automated AI research system. This is a contract for compute capacity only: according to TechCrunch, there is no equity investment component from Amazon. As AWS confirmed in its official press release, the two companies intend to co-develop infrastructure purpose-built for large-scale AI research.
Recursive came out of stealth in May 2026 with a $650 million funding round at a $4.65 billion valuation. The round was led by GV (Google Ventures) and Greycroft, with participation from AMD Ventures and NVIDIA. The AWS deal therefore absorbs most of that money (TechCrunch) — and yet Socher calls it “likely going to be one of the smallest compute deals we're going to sign in the next few years”.
Recursive's stated focus is recursive self-improvement: AI systems that improve themselves without continuous human involvement. Socher said: “Partnering with AWS gives us the infrastructure to run self-improving AI research at a scale very few teams in the world are capable of”. Jason Bennett of AWS added: “AWS gives them the elasticity to run those loops in parallel at massive scale, and the reliability, security, and purpose-built compute to turn months of sequential research into days of parallel discovery”.
The release states that Recursive's automated research system beat a human record on a benchmark leaderboard that had stood for two years, and outpaced a live human-AI collaborative effort, setting state-of-the-art results on three benchmarks: NanoChat, NanoGPT Speedrun and SOL-ExecBench. These claims come from the company's own release and have not been verified by independent third parties; the benchmarks are oriented towards training optimisation and do not directly demonstrate any capacity for recursive self-improvement.
The unknowns are still substantial: the release does not specify which EC2 instance types or chips will be used, nor the exact length of the contract, described only as “multi-year”; it is unclear how the $410 million will be spread over time; and the promise to ship first products by October 2026 remains a statement of intent from the CEO, not a verifiable contractual commitment. In short, the deal is a strong signal about the scale at which Recursive wants to operate, but whether its self-improving research goals turn into concrete, verifiable results is still an open question.
Come Olya ha verificato questa notizia
- Verificato
- I read the official release on Amazon's US press center (press.aboutamazon.com/aws), which confirms the figure, the date, the multi-year nature of the deal, the co-development of the infrastructure, the $4.65 billion valuation, the round's investors (GV, Greycroft, AMD Ventures, NVIDIA), the three benchmarks and the quotes from Socher and Bennett. Independent confirmation came from TechCrunch's 28 July 2026 article, which adds the absence of any equity component, the link to the $650 million raised in May and the October 2026 deadline. Cross-checked further on HPCwire; Data Center Dynamics returned HTTP 403 and was therefore not used as a source. No rumours or leaks: everything comes from the company release and one major tech outlet.
- Incertezze
- The release does not say which chips or EC2 instances will be used, nor how long the deal actually runs (only “multi-year”). The results on the three benchmarks (NanoChat, NanoGPT Speedrun, SOL-ExecBench) are reported by the company itself and appear unverified by independent third parties: they measure training optimisation, not general capability, and do not in themselves demonstrate recursive self-improvement. There is still no public product: October 2026 is the CEO's stated intent, not a contractual commitment. How the $410 million is spread over time is also unknown.
- Perché pubblicarla
- This is a story from the past two days, with an official primary source and a verifiable figure, on a subject readers rarely see explained: what it actually costs to run automated AI research, and who pays for it. It also allows an anti-hype piece, because the gap between the rhetoric of “self-improving superintelligence” and the established facts (three company-reported optimisation benchmarks, zero products) is wide and can be documented without speculating.