[ AGI & Future AI
](https://www.unite.ai/series/agi/)
[Add Unite.AI to your preferred sources on Google](https://www.google.com/preferences/source?q=unite.ai)
Recursive, the AI research company led by Richard Socher, has signed a multi-year, $410 million agreement with Amazon Web Services (AMZN ) to run its automated AI research system, putting the bulk of its funding into a single cloud contract.
The company came out of stealth in May 2026 with $650 million at a $4.65 billion valuation, which makes the AWS commitment close to two-thirds of everything it has raised. Recursive’s own site puts the team at more than 25 people across San Francisco and London. That ratio is the substance of the deal: the company is buying compute on frontier-lab terms while staffing like a seed-stage startup.
Socher, Recursive’s CEO and co-founder, told TechCrunch the 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 workload behind the spend #
Recursive’s system runs the research loop itself: propose a change, implement it, run the experiment, validate the result, then choose the next experiment. Its first public results, published June 11, 2026, show what that consumes in practice. The three benchmarks the company reported are individually tiny: a five-minute language-model training budget on a single GPU, a training speedrun scored on one eight-GPU H100 node, and 235 GPU kernel-writing tasks measured on Nvidia (NVDA ) B200s.
On the speedrun, Recursive reported cutting the time to a fixed validation loss from 79.7 seconds to 77.5 seconds, against a leaderboard the open-source community had been optimizing for more than two years. On the kernel benchmark, it reported lifting the mean score from 0.699 to 0.754 on a scale where 1.0 represents the analytical hardware limit.
Small jobs, run thousands of times, are what make this expensive. Each experiment spawns the next, and every promising candidate needs a validation pass to screen out solutions that game the evaluator instead of improving the model. That is a different procurement problem from a months-long pretraining run on a dedicated cluster. It rewards burst capacity, fast scheduling and high job churn over one enormous reserved block of accelerators.
AWS is selling against that profile. Jason Bennett, its vice president and global head of startups and venture capital, said in the announcement that self-improving AI creates compounding compute demand because every research loop generates the next experiment, and pitched AWS on the elasticity to run those loops in parallel. The two companies also said they plan to co-develop infrastructure built specifically for large-scale automated research, a commitment that outlasts any single block of capacity.
There is a second-order effect worth noting: much of what Recursive’s system has produced so far is efficiency. Faster kernels and cheaper training recipes are the output, which means part of the AWS bill funds work aimed at lowering the cost of the compute that follows.
Priced in dollars, not gigawatts #
AWS’s largest AI commitments are denominated in power rather than currency. In its first-quarter results on April 29, 2026, Amazon said OpenAI had committed to consume roughly two gigawatts of Trainium capacity through AWS infrastructure, ramping from 2027, and that Anthropic would secure up to five gigawatts of current and future generations of the same chips. AWS revenue in that quarter was $37.6 billion, up 28% year over year, and Amazon’s custom silicon business passed a $20 billion annual run rate.
Set beside those, $410 million spread across several years is a small line for AWS. It is also structured differently. TechCrunch reported that Amazon is taking no investment stake alongside the capacity, which separates this from the compute-and-capital packages that have shaped recent frontier-lab financing, from AMD’s $5 billion Anthropic bet to Anthropic’s takeover of SpaceX’s Colossus 1. Recursive is buying capacity as a customer, so the contract sits on its own balance sheet rather than inside a strategic partnership.
That is the harder version of the trade, and it explains the ratio. A company automating its own research process converts capital into experiments rather than salaries, and the cost of running a research organization shifts from payroll onto the cloud bill.
What ships next #
Recursive’s stated sequence starts with AI research and widens from there: use the automated loop to improve AI systems first, then apply the same method to other scientific problems. The compute is now contracted for years, and the co-development work with AWS gives the company input on the systems its experiments run on.
The nearer milestone is commercial. Socher said in the same interview that the first products people can use should appear around October 2026, and that they will arrive in months rather than quarters.