Mirendil has signed a multi-year Google Cloud deal worth upwards of $100M to train its self-improving AI systems, co-founder and CEO Behnam Neyshabur confirmed. The commitment equals roughly half of the seed funding the startup closed in late June at a $1B valuation, an aggressive compute bet for a company at seed stage.
The agreement gives Mirendil access to Google TPUs, Nvidia GPUs, and managed training clusters. That mixed-silicon footprint is the point: the startup's research program hinges on assigning distinct workloads to the accelerator best suited for each, and Google is one of the few providers that can offer both first-party and Nvidia hardware inside a single stack.
Neyshabur and co-founder Harsh Mehta are both Anthropic alumni, and Mirendil's pitch sits squarely in the recursive self-improvement lane that Anthropic and others have been exploring. The idea is an AI system that iteratively improves its own capabilities, accumulating expertise the way a human researcher does across a career. Neyshabur wants that loop pointed at hard scientific problems.
“You can have a self-improving AI where you can point a problem at it and it keeps getting better with time”— Behnam Neyshabur, Mirendil co-founder and CEO
Key facts
- 01Mirendil signed a multi-year Google Cloud deal worth upwards of $100M for TPUs, Nvidia GPUs, and managed training clusters.
- 02The commitment equals roughly half of the seed round Mirendil closed in late June at a $1B valuation.
- 03Co-founders Behnam Neyshabur and Harsh Mehta previously worked at Anthropic, joining a wave of startups chasing recursive self-improvement.
- 04Mirendil aims to build an AI that can do the work of an entire frontier lab, targeting research in medicine, biology, and materials science.
- 05Rival startups Recursive Superintelligence and Ricursive Intelligence are pursuing similar recursive self-improvement goals.
"How can we have an AI system that keeps doing research, keeps improving its own knowledge and performance when it comes to Alzheimer's disease?" Neyshabur said. He frames the ambition as automating the work of an entire frontier lab, with medicine, biology, and materials science as the initial targets.
Two other startups, Recursive Superintelligence and Ricursive Intelligence, have emerged around the same thesis, though none of the three has yet demonstrated a production system doing what the phrase promises. The category is early enough that most of the investment is going into the compute bill.
“These models are really good at working with different workloads and chips, and assigning the right workloads to the right chips”— Harsh Mehta, Mirendil co-founder
That is where the Google deal matters. Training a system designed to rewrite its own training loop is compute-heavy in a way that scales worse than standard pretraining, because each self-improvement cycle can trigger a new run. Mehta said Mirendil's software layer is built to route each phase of that cycle to the cheapest suitable chip in Google's fleet — a claim Google will happily reinforce because it needs a differentiation story against AWS and Azure that runs deeper than raw FLOPs.
Mehta added that Google's willingness to provide multiple kinds of chips lets Mirendil "mix and match workloads with the right kind of accelerators, and then lower the cost not just for us, but also for our customers using our systems." The subtext is that Mirendil intends to resell orchestration on top of Google's hardware, not just consume it.
Amin Vahdat, Google's SVP and chief technologist of AI and infrastructure, framed the partnership as validation of Google's systems-level approach to AI compute. For Google, the trade is straightforward: a nine-figure compute commitment from a seed-stage lab, plus a strategic reference customer building frontier recursive-self-improvement technology that Google can eventually pitch to enterprise buyers.
The structure of the deal — cloud provider fronts massive compute capacity to a young AI lab in exchange for revenue commitments and strategic alignment — has become the dominant pattern of 2026. Microsoft did versions of it with OpenAI, Amazon and Google both did it with Anthropic, and the smaller variants are proliferating as cloud giants try to lock in the next tier of labs before the market consolidates.
The open question is whether recursive self-improvement is a real research program or a marketing frame stretched over conventional model training. Nothing about Mirendil's technical approach has been published in detail, and the phrase has been used loosely across the industry. A $100M compute deal buys the runway to find out; it does not answer the question.
For Google Cloud, Mirendil is a useful data point in a longer argument. The company has spent 2026 trying to convince the market that its combination of TPUs, GPUs, and orchestration software is a viable alternative to a pure-Nvidia stack, and every frontier-adjacent lab it signs on those terms tightens that argument. If Mirendil's self-improving systems actually produce research output, Google gets a showcase customer inside a category the hyperscalers are all watching.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
