A London-based startup developing continually self-improving AI has agreed to a $410 million compute deal with AWS.

Recursive Superintelligence, which emerged from stealth in May with a valuation of $4.65 billion, has raised $650 million in investment to date and said the AWS agreement has consumed most of its funding so far.

At this stage, a more traditional move for a startup would be to expand its staff. Recursive, however, says its strategy is more appropriate for a company that is focused on automating the development of its own product.

The startup is building AI that performs its own experiments on how to improve safely, in an open-ended process of ongoing evolution. It therefore says “agent count” matters more than headcount.

Recursive argues that the ability to keep upgrading itself can reduce the need for human input until a closed loop is formed, leaving the system limited primarily by the amount of computing power it can access. The AWS arrangement is central to that model, and the size of the deal underscores how compute access has become a core strategic input for frontier AI developers as training and experimentation demands grow.

“Self-improving AI creates a compounding demand for compute -- every research loop generates the next experiment. In just a few months, Recursive has produced results that outpaced years of community-driven optimization,” Jason Bennett, vice president, worldwide startups and venture capital at AWS, said in a statement. “AWS gives them the elasticity to run those loops in parallel at massive scale, and the reliability, security, and purpose-built compute to turn months of sequential research into days of parallel discovery.”

The results Bennett referred to were published by Recursive and highlighted leading performance in fixed-budget large language model training, small-model training speed, and GPU kernel optimization.

The company said that progress has also led AWS and Recursive to co-develop infrastructure designed for large-scale automated AI research, which could be useful for other companies focused on similar work in the future.

“Recursive self-improving AI represents one of the biggest shifts ahead for humanity. The clearest path to superintelligence is to build AI that recursively improves itself,” co-founder and CEO Richard Socher said in a statement. “AI is code, and AI can code. If we can close that loop properly, we can automate the scientific method itself -- the same process that has driven humanity’s progress in discovery and invention for centuries.”

Financial backers of the startup include GV (Google Ventures), Greycroft, AMD Ventures and Nvidia.