AI startup Recursive Superintelligence strikes $410 million compute deal with Amazon
A company built around AI that improves itself has committed most of its funding to Amazon Web Services, betting that raw computing power is the real bottleneck.
Key points
- Recursive Superintelligence signed a multi-year compute deal with Amazon Web Services worth $400 million, first reported by TechCrunch AI.
- The startup emerged from stealth in May 2025 with $650 million in total funding.
- The $400 million commitment represents the bulk of the company's money raised to date.
- CEO Richard Socher expects this deal to be among the company's smallest future compute agreements.
- Recursive plans to release its first public products around October 2025.
A little-known AI startup has just committed most of its war chest to raw computing power. Recursive Superintelligence, which only announced its existence in May, signed a multi-year agreement on Tuesday to spend $400 million on compute capacity through Amazon Web Services, the cloud computing arm of Amazon.
Compute, in this context, means the specialised chips and server time that AI systems need to run and train. The more ambitious the AI, the more compute it typically burns through.
Why spend so much on computing instead of people?
Because Recursive's core bet is that AI can improve itself, rather than waiting for human engineers to do the work.
The company focuses on recursive self-improvement, or RSI: the idea that an AI system can analyse its own weaknesses and redesign itself to fix them, without constant human direction. If that works, a small team could produce output normally requiring hundreds of engineers. The catch is that the process is extraordinarily demanding on computing resources.
"For us, it's less about headcount and more about agent count," CEO Richard Socher said. An agent, here, means software that carries out multi-step tasks on its own, without someone clicking through each step.
Socher told TechCrunch AI the $400 million deal would likely be "one of the smallest compute deals we're going to sign in the next few years," suggesting the company expects its appetite for processing power to grow sharply. That framing echoes what we saw in our 14 July report on Reflection AI's $1 billion compute deal with Nebius: large upfront commitments are becoming a standard opening move for frontier AI labs, not a sign of distress.
What does Amazon get out of this?
No equity stake. Amazon is purely a vendor here, unlike the bundled investment arrangements some major labs have struck with cloud providers.
What AWS does get is a high-profile customer with unusual technical requirements. Jason Bennet, Vice President for startups and venture capital at AWS, said the agreement includes co-developing infrastructure designed specifically for self-improving AI systems. That custom work could give Amazon a template to attract other frontier AI companies.
Recursive's total funding stands at $650 million. Spending $400 million of that with a single vendor is aggressive, and it signals how central raw computing capacity is to the company's strategy.
Should ordinary people care about self-improving AI?
Not urgently, but this technology is moving from theory toward product.
Self-improving AI has long been discussed as a potential turning point: once AI can redesign itself, progress could accelerate faster than humans can track. Researchers disagree sharply on whether that moment is close or distant, and whether it looks like a smooth improvement or a sudden leap.
Recursive isn't waiting for the debate to settle. Socher said the goal is practical products ordinary users can try. "You will see those within a few months, not within a few quarters or years," he said, with October as the target for the first public release.
The thing to watch isn't the announcement. It's whether those October products show genuine self-improvement in action, or whether they arrive looking like conventional AI tools dressed in a bold origin story.
