Recursive Superintelligence, Richard Socher's startup building self-improving AI systems, signed a multi-year agreement with Amazon Web Services on July 28, 2026 for $410 million (approx. PLN 1.64 billion) of compute?compute: the computing power (GPUs/TPUs) needed to train and run AI models. It represents the majority of the funding the company raised in May 2026 — and, according to its founder, one of the smallest compute deals it expects to sign.
Key takeaways
- AWS deal: $410 million (approx. PLN 1.64 billion), multi-year, with no equity stake from Amazon
- Recursive Superintelligence emerged from stealth in May 2026 with $650 million and a $4.65 billion valuation
- Round backers included GV, Greycroft, Nvidia and AMD
- AWS will co-develop infrastructure purpose-built for this type of company
- First tangible products are expected in October 2026
Compute over headcount
Richard Socher summed up the company's logic in one line: what matters is not headcount but agent count. Instead of scaling a team of people, Recursive Superintelligence bets on scaling a fleet of AI agents, and that requires compute above all. Hence the unusual split: most of the raised capital goes not to salaries but to compute from Amazon's cloud.
The founder described the current deal as likely “one of the smallest we're going to sign” in the coming years. That signals the company expects a sharp rise in compute demand as its systems mature.
What Recursive Superintelligence is building
The company was founded to build systems capable of recursive self-improvement — AI that iteratively improves itself. Recursive Superintelligence emerged from stealth in May 2026 with $650 million (approx. PLN 2.6 billion) in funding and a $4.65 billion (approx. PLN 18.6 billion) valuation. As The New York Times reported, notable researchers joined the effort, and the round was backed by GV, Greycroft, Nvidia and AMD.
The presence of Nvidia and AMD among the investors is telling: both chipmakers have a direct interest in firms that buy vast amounts of compute. The compute deal itself, however, is with Amazon Web Services, not with the accelerator makers.
A deal, not an investment
Unlike many high-profile deals of recent months, in which a cloud provider took a stake in an AI customer, Recursive's agreement with Amazon has no investment component. It is a pure compute-services contract, supplemented by co-developed infrastructure.
We're going to co-develop infrastructure purpose-built for these types of company.
Jason Bennett, VP of Startups and Venture Capital at AWS.
Why it matters
The deal shows how the cost model of agent-first AI companies is changing. When the main “worker” becomes an agent rather than a person, the biggest budget line stops being payroll and becomes the compute bill. The claim that $410 million is one of the smallest future deals hints at the scale of spending the company deems necessary to train and run self-improving systems.
For Amazon, in turn, it is a way to lock in a promising customer without putting up equity — a model that lets it compete for the most compute-hungry startups. Caution is warranted, though: the company has not shown a working product yet, and recursive self-improvement remains a research goal, not a proven production technology.
What's next?
- First “tangible, useful” products to try are promised for October 2026, according to the founder
- Larger compute deals are explicitly flagged as part of the company's strategy
- The AWS co-developed purpose-built infrastructure will test whether cloud providers can tailor hardware to agent-first firms
