Richard Socher's Recursive Superintelligence Locks In $410M AWS Compute Deal

Recursive Superintelligence has committed roughly two-thirds of the $650 million it has raised to date to a single line item: compute. The London-based startup, led by former Salesforce chief scientist Richard Socher, has signed a multiyear deal worth $410 million with Amazon Web Services to power its research into self-improving AI systems.

The arrangement is structured purely as a compute purchase. AWS is not taking an equity stake as part of the agreement, a distinction that separates this deal from the increasingly common practice of cloud providers trading credits for ownership in frontier AI labs.

Beyond simply renting GPUs, AWS and Recursive say they will jointly build infrastructure tailored to large-scale automated AI research, the kind of workload that requires running thousands of experiments in parallel rather than training a single monolithic model. That focus reflects Recursive's core thesis: rather than shipping one large language model, the company is trying to automate the process of AI research itself.

Recursive has only been public since May, when it exited stealth carrying a $4.65 billion valuation, an unusually high figure for a company with no commercial product yet on the market. Its investor roster, which includes GV, Greycroft, AMD Ventures and Nvidia, signals interest from both traditional venture capital and chipmakers with a direct stake in compute demand.

What the early results show

The company has pointed to published benchmark results as evidence its approach works. It claims leading performance in fixed-budget large language model training, in accelerating small-model training speed, and in GPU kernel optimization, three areas where incremental efficiency gains translate directly into lower compute costs at scale.

Recursive says it wants to show early product examples as soon as October, a tight timeline for a company still defining what a commercial offering built on self-improving AI research would even look like. That deadline will be the first real test of whether the compute investment converts into something customers can use.

The bet underlying the AWS partnership is that automating AI research, rather than just scaling a single model, is where the next efficiency gains will come from. Whether that thesis holds will depend less on benchmark papers and more on what ships this fall.