Key Takeaways:
Recursive, the AI research startup led by Richard Socher, committed $410 million to Amazon Web Services for cloud compute capacity — roughly two-thirds of its total raised capital — in a deal that shows how automated AI research consumes infrastructure at hyperscaler scale.
Amazon Web Services signed a $410 million multi-year agreement with Recursive, an AI research startup automating scientific discovery, to run the company's self-improving research system on AWS infrastructure.
"Self-improving AI creates a compounding demand for compute — every research loop generates the next experiment," Jason Bennett, vice president and global head of startups and venture capital at AWS, said in a statement.
The deal consumes close to two-thirds of the $650 million Recursive raised when it emerged from stealth in May at a $4.65 billion valuation, led by GV and Greycroft with participation from AMD Ventures and Nvidia. The company, with just over 25 employees across San Francisco and London, plans to use the capacity to run thousands of small-scale experiments that improve its AI models autonomously — a workload profile that favors burst capacity and rapid job scheduling over dedicated clusters.
The contract positions AWS to capture demand from a new class of AI customer: labs that convert capital into experiments rather than headcount. Amazon's cloud unit reported $37.6 billion in revenue in the first quarter, up 28% year over year, and its custom silicon business has passed a $20 billion annual run rate. Recursive's first products are expected around October, Socher said.
The Compute Profile Behind the Deal
Recursive's system runs a closed research loop: propose a change, implement it, run the experiment, validate the result, then choose the next experiment. Its first public results, published June 11, showed the system beating a two-year human leaderboard record on the NanoGPT Speedrun benchmark — cutting time to a fixed validation loss from 79.7 seconds to 77.5 seconds — and improving scores on a GPU kernel-writing benchmark from 0.699 to 0.754 on a scale where 1.0 represents the hardware limit.
Each experiment is small, but the loop generates thousands of them. That procurement profile differs from the months-long pretraining runs that dominate frontier-lab spending at OpenAI and Anthropic. AWS is selling elasticity against that need, and the two companies plan to co-develop infrastructure built specifically for large-scale automated research.
Competitive Context and Investor Impact
The deal is structured as a pure customer contract with no equity stake, separating it from the compute-and-capital packages that have shaped recent frontier-lab financing. AMD committed $5 billion to Anthropic, and Anthropic took over SpaceX's Colossus 1 data center. Recursive is buying capacity on its own balance sheet.
Socher told TechCrunch the $410 million agreement is "likely to be among the smallest" the company signs over the next few years. "For us, it's less about headcount and more about agent count," he said.
The deal is small relative to AWS's largest AI commitments. OpenAI committed roughly two gigawatts of Trainium capacity through AWS, ramping from 2027, and Anthropic secured up to five gigawatts. But the Recursive contract signals a growing segment of AI infrastructure demand: research automation that compounds compute consumption with every experimental cycle.
Amazon shares trade at roughly 22 times forward earnings. The company's custom silicon business, now at a $20 billion annual run rate, gives AWS a cost advantage in serving AI workloads that competitors using Nvidia's merchant silicon cannot fully match. For Nvidia, which supplies the B200 chips Recursive used in its kernel benchmark, the deal represents a customer that may eventually design around its hardware as AWS's Trainium ecosystem matures.
This article is for informational purposes only and does not constitute investment advice.
