Richard Socher, the former Salesforce chief scientist who co-founded Recursive, announced Tuesday that his company had signed a multi-year agreement worth $410 million with Amazon Web Services. Under the deal, Recursive will designate AWS as its cloud provider and run its automated AI research system on the platform, which the startup describes as an engine for building AI that improves itself. The two companies will also co-develop infrastructure purpose-built for large-scale automated research.
Recursive is one of a handful of startups pursuing recursive self-improvement, the idea that an AI system can propose ideas, run experiments, validate results and use what it learns to design the next experiment, closing the research loop without a human in the middle. The company frames the approach as automating the scientific method itself. Its founding team includes former research leaders from OpenAI, Google DeepMind, Meta AI, Salesforce AI and Uber AI, spanning natural language processing, reinforcement learning, robotics and computer vision.
The startup’s first public results, released this month, give the pitch some empirical footing. Recursive said its automated system beat a two-year human leaderboard record and outperformed a live collaborative human-AI effort, setting state-of-the-art results across three benchmarks: NanoChat, NanoGPT Speedrun and SOL-ExecBench. The claim, if it holds up under replication, has an unusual implication: compute budget, rather than headcount, becomes the primary scaling lever for research progress. That helps explain why $410 million of the roughly $650 million the company has raised is headed to cloud infrastructure rather than salaries.
AWS’s pitch for the business was about parallelism. Jason Bennett, vice president and global head of startups and venture capital at AWS, said self-improving AI creates compounding demand for compute, because every research loop generates the next experiment. In just a few months, he said, Recursive has produced results that outpaced years of community-driven optimization. AWS gives the company the elasticity to run those loops in parallel at massive scale, and the reliability and purpose-built compute to turn months of sequential research into days of parallel discovery.
For Amazon, the deal is a small but symbolic piece of a much larger competition. AWS, Microsoft and Google are spending aggressively to keep frontier labs and ambitious startups on their clouds, and compute agreements in the hundreds of millions have become a standard instrument in that fight. Recursive’s decision to name AWS its provider, and to co-develop infrastructure with the unit, suggests the startup expects its compute needs to grow well beyond this agreement. Socher was explicit about the trajectory, telling reporters the $410 million will likely be one of the smallest compute deals the company signs in the next few years.
The arrangement also carries a strategic hedge for Recursive. Multi-year cloud commitments of this size typically lock in capacity and pricing, protecting a young company from the shortages and cost spikes that have hit AI labs during training runs. The co-development clause goes further, effectively turning AWS engineers into collaborators on the specialized research infrastructure Recursive says it needs. For AWS, that same clause is a way to deepen the relationship and raise the cost of switching to a rival cloud later.
The arrangement also fits a pattern in the cloud market that has accelerated since the start of the year. OpenAI’s agreement to buy compute from Microsoft, Anthropic’s commitments across multiple clouds and a string of hundred-million-dollar deals between startups and hyperscalers have made compute commitments a standard part of the funding stack for AI companies. In that crowd, a $410 million commitment is large but not exceptional, and Recursive’s executives said as much themselves.
Not everyone is convinced the underlying thesis will hold. Automated research systems have a history of impressive results on narrow benchmarks and much less to show at the frontier, where problems are ill-defined and failures are expensive. Skeptics note that the benchmarks Recursive cites are small enough that a well-designed search procedure might outperform humans without generalizing to harder domains. The company’s own rhetoric does not help: Socher has described the goal as nothing less than automating knowledge discovery, a claim that invites scrutiny every time a result is released.
The broader context is an industry awash in infrastructure money. Hyperscaler capital spending and startup compute agreements have pushed more capital into raw compute in 2026 than in any prior year, and cloud providers are increasingly comfortable funding the demand they hope to serve. Deals like this one shift some of the risk of that buildout onto the cloud giants, who recover it through usage fees if the startup’s research consumes what it promises.
For Recursive, the agreement buys time, capacity and credibility. For AWS, it is a bet on a thesis: that the next generation of AI research will be measured in megawatts as much as in models. If automated research matures, this $410 million contract will look like a cheap seat at the table. If it does not, it will look like one more generous cloud subsidy for a startup with a big idea and no customers. The market will not know which for several years, but the checks are already being written.
