Mirendil inks $100M+ Google Cloud deal to build AI that recursively improves itself

Mirendil, a startup barely two months past a $200 million seed raise, just locked in a compute deal with Google Cloud worth north of $100 million. That’s half the company’s entire seed funding, spent on raw infrastructure before it has publicly shipped a product. The multiyear pact gives Mirendil access to both Google’s TPUs and Nvidia GPUs, plus managed training clusters — the kind of horsepower you need for recursive self-improvement, the lab’s entire raison d’être.

The bet, as CEO Behnam Neyshabur describes it, is on AI that doesn’t just answer prompts but iteratively sharpens its own capabilities over time. Point it at Alzheimer’s research, he told TechCrunch, and it keeps getting better at the problem without humans manually retraining it. It’s a pitch that’s attracting serious talent — Neyshabur and his co-founders came out of Anthropic, and a small crop of rival startups like Recursive Superintelligence are chasing the same grail.

But recursive self-improvement is a compute furnace. Co-founder Harsh Mehta framed the Google deal as a workload-matching play: different AI tasks run on different chips, and Mirendil’s software layer routes the right jobs to the right silicon. That flexibility, he argues, cuts costs for the startup and eventually for customers using its systems. Google gets something juicier in return — a frontier lab building next-generation AI entirely on its infrastructure stack, which becomes a showcase for enterprise clients weighing cloud providers.

Amin Vahdat, Google’s AI infrastructure chief, said the industry has moved past obsessing over chip-level specs to “how we orchestrate entire systems of intelligence.” That’s the cloud giant’s strategic dodge against AWS and Microsoft: lock in the most ambitious labs early, and let their workloads prove the platform. For Mirendil, the clock is now ticking. A $100 million compute tab before revenue means the engineering had better deliver something that actually self-improves — not just a whitepaper about it.