Mirendil, a startup founded by Anthropic alumni, has committed to a multi-year compute agreement with Google Cloud worth more than $100 million, TechCrunch reported exclusively on August 6, citing chief executive Benham Neyshabur. The commitment lands about two months after Mirendil closed a seed round that valued the company at $1 billion, and the new cloud spend equals close to half of that raise.
The arrangement gives Mirendil access to Google’s tensor processing units alongside Nvidia’s graphics chips, plus training infrastructure Google will manage on the startup’s behalf. Mirendil intends to use that capacity to pursue what researchers call recursive self-improvement: systems built to refine their own performance without a human redesigning them at every step. Neyshabur described the goal as building AI that keeps advancing on a single problem once it is set loose on one, pointing to Alzheimer’s research as a domain where sustained, self-directed progress would matter most.
That ambition sits inside a crowded, young field. Anthropic, where several Mirendil founders previously worked, has its own recursive self-improvement research effort underway. Rivals including Recursive Superintelligence, which TechCrunch separately reported signed a compute deal with Amazon worth roughly $400 million in late July, and a smaller outfit called Ricursive Intelligence are chasing the same technical bet with different cloud backers.
That pattern points to something the headline dollar figure alone does not capture: compute contracts at this scale now function as a form of vendor financed distribution. Cloud providers are not simply selling capacity. They are underwriting a startup’s entire research roadmap in exchange for becoming the infrastructure that startup’s eventual product runs on, and later resells to enterprise buyers. Google’s win here matters competitively because Amazon already claimed Recursive Superintelligence as a tenant, and Microsoft’s Azure unit has been courting its own frontier-research customers. Landing a lab built by Anthropic veterans gives Google a reference account it can point to when pitching TPU capacity against Nvidia-centric cloud rivals.
The math also says something about Mirendil’s spending posture. A nine-figure compute commitment against a seed round of roughly $200 million means the company is directing a large share of its early capital toward infrastructure before it has a shipping product or disclosed revenue. That is typical for a frontier-style research lab, but it leaves little room for a slow ramp. Burn at this scale needs either a fast path to a sellable capability, or a follow-on round sized closer to what Anthropic and OpenAI already spend on compute.
Google frames the deal as validation of a pitch built around mixing chip types rather than competing purely on TPU benchmarks. Amin Vahdat, the company’s chief technologist for AI and infrastructure, said progress now depends on coordinating whole systems of models and hardware, not any single chip’s speed. Mirendil co-founder Harsh Mehta echoed that framing, saying the startup’s software assigns each workload to whichever accelerator suits it best, an approach meant to cut costs for Mirendil and for the customers eventually using its systems.
Operators evaluating their own cloud commitments should watch whether Mirendil produces a research result within the next year, not just a funding headline. A compute deal this size only proves the underlying thesis if it converts into something Google, or a paying customer, can point to.
Reported exclusively by TechCrunch on August 6, 2026.
