Mirendil signs $100M Google Cloud deal to train self-improving AI

Mirendil, an AI lab founded by former Anthropic researchers, has agreed to a multi-year Google Cloud partnership worth more than $100 million. 

Its CEO, Behnam Neyshabur, confirmed the partnership and stated that it was to secure the compute behind its self-improving AI.

Mirendil raised $200 million in a seed round at a $1 billion valuation in late June. It is still considered one of the largest seed rounds for an AI startup, and it was led by Andreessen Horowitz with Kleiner Perkins and Nvidia among the backers.

Neyshabur put the Google Cloud commitment at upwards of $100 million, which works out to roughly half of that seed capital going toward a single infrastructure deal.

The agreement opens access to Google’s TPUs and Nvidia GPUs. It also comes with managed training clusters.

Mirendil’s cofounder Harsh Mehta stated, “These models are really good at working with different workloads and chips, and assigning the right workloads to the right chips,” adding that the flexibility lowers cost “not just for us, but also for our customers using our systems.”

What does “self-improving AI” actually mean?

Self-improving AI, or recursive self-improvement, refers to systems that iteratively rework themselves.

Mirendil’s ambition is to build a model that will be able to do the job of a whole frontier lab.

Its CEO, Neyshabur, has worked at Anthropic and Google as a research scientist, leading teams in various projects. He is also a co-inventor of the Sharpness-Aware Minimization (SAM) optimization algorithm.

On their ambition, Neyshabur said, “You can have a self-improving AI where you can point a problem at it, and it keeps getting better with time,” citing Alzheimer’s research as the kind of open-ended goal the company wants its system to chase.

When he was at Anthropic, cofounder Mehta contributed to internal efforts to automate parts of its own research.

Mirendil says its models are meant to speed up work in medicine, biology, chemistry, and materials science.

Anthropic is also working on the trend, as it reported that more than 80% of the code it merged as of May 2026 was written by its Claude models, up from low single digits before early 2025, and that a typical engineer now merges eight times as much code per day as in 2024.

However, Anthropic mentioned that the metric overstates real productivity gains but calls the acceleration genuine.

Why Google wanted the deal

For Google, the partnership is a bet on a customer that could sell the cloud giant’s own hardware. Neyshabur stated that Mirendil’s software layer helps customers wring more out of Google’s chips, which gives Google a pitch against rival clouds.

What Google gets in return is a strategic partner that is building frontier recursive self-improving AI, and this is a technology it can later market to enterprise buyers.

Amin Vahdat, Google’s SVP and chief technologist for AI and infrastructure, said that progress is no longer just about chip-level performance “but how we orchestrate entire systems of intelligence and break through the physical constraints of scaling.”

The deal also highlights an ongoing scramble among infrastructure providers, in this case, cloud providers.

Mirendil is not alone in chasing recursive self-improvement either, as there are newer entrants, such as Recursive Superintelligence and Recursive Intelligence, who are pursuing the same goal.