Exclusive: Mirendil inks $100M+ Google Cloud deal to scale self-improving AI

Mirendil, the AI lab founded by ex-Anthropic researchers, has struck a multi-year compute deal with Google Cloud worth more than $100 million, according to CEO Benham Neyshabur. That's roughly half of what the company raised in its seed round back in June, when it landed a $1 billion valuation without shipping a product the public has heard of. The math alone tells you something about where AI startup priorities sit right now: raise big, then immediately hand a huge chunk of it to a cloud provider before you've proven much of anything.

What Mirendil is actually building is recursive self-improvement — AI systems designed to get better at their own job over time, without a human rewriting the code each cycle. Neyshabur frames it less like a chatbot upgrade and more like mimicking a scientist's career: you start clumsy in a field, accumulate expertise, and get sharper the longer you work a problem. His pitch example is blunt. Point the system at Alzheimer's research and let it keep grinding, indefinitely, getting more capable as it goes. That's the promise, anyway. Whether it holds up outside a demo is the open question nobody in this space has answered yet.

The deal gives Mirendil access to both Google's TPUs and Nvidia GPUs alongside managed training clusters, which matters because self-improving AI eats compute at a pace regular model training doesn't. Co-founder Harsh Mehta says a lot of the real engineering work is matching specific workloads to specific chips rather than just throwing more silicon at the problem — squeezing efficiency out of a mixed hardware fleet instead of betting everything on one accelerator type. Google, for its part, is happy to be the flexible option; SVP Amin Vahdat talked about orchestrating