Mirendil has made a major compute move as it pursues self-improving AI. The AI lab has signed a multi-year partnership with Google Cloud, a deal that co-founder and CEO Benham Neyshabur told TechCrunch is worth upwards of $100 million.
The agreement gives Mirendil access to Google’s TPUs, Nvidia GPUs and managed training clusters. For a startup trying to build AI systems that can keep improving through research and iteration, that infrastructure is not a side issue. It is central to the work.
A compute deal built for self-improving AI
Mirendil is working on self-improving AI, also known as recursive self-improvement. The idea is to build AI systems that can iteratively improve themselves, gaining knowledge and performance over time as they work on a problem.
The company’s ambition is large. Mirendil hopes its AI will eventually be able to take on the work of an entire frontier AI lab. That means the system would not simply answer questions or perform narrow tasks. It would work on research itself, making progress toward goals that require accumulated knowledge and continued improvement.
Neyshabur described the vision in practical terms: “You can have a self-improving AI where you can point a problem at it and it keeps getting better with time,” he said.
He also framed the approach around scientific questions. “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?” he continued. “This technology allows us to set goals that are ambitious for AI, and the AI would keep making progress.”
Why the Google Cloud partnership is significant
The deal is worth upwards of $100 million, roughly half of what Mirendil raised in seed funding at a $1 billion valuation in late June. That scale shows how expensive the infrastructure side of frontier AI research has become, especially for companies focused on training systems that may need to run many kinds of workloads.
The partnership also reflects a broader shift in the AI market. Cloud giants are pursuing startups with large infrastructure commitments, while AI companies are trying to lock in as much compute as possible before they need it. In that environment, access to hardware can shape what research a company can attempt and how quickly it can scale.
For Mirendil, the value is not only raw capacity. The agreement includes access to multiple kinds of accelerators and managed training clusters. That matters because training self-improving AI requires enormous amounts of computing power, and different workloads may perform better on different hardware.
The hardware mix behind the research
Mirendil’s co-founder Harsh Mehta said training increasingly depends on matching the right work to the right chips. The company sees flexibility across Google’s TPUs and Nvidia GPUs as a way to improve how workloads are assigned and executed.
“These models are really good at working with different workloads and chips, and assigning the right workloads to the right chips,” Mehta said. “[Google] provides multiple kinds of chips […] this flexibility allows us to ultimately 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.”
That point is important because AI infrastructure is no longer only a question of having the fastest individual chip. Mirendil’s research depends on systems that can coordinate hardware, software and workloads at scale. If the system can place the right computation on the right accelerator, the same cloud partnership can support a wider range of research needs.
Google is presenting that flexibility as part of its infrastructure strategy. Amin Vahdat, SVP and chief technologist of AI and infrastructure at Google, said in a statement that AI advancement is not only about chip-level performance anymore, “but how we orchestrate entire systems of intelligence and break through the physical constraints of scaling.”
What Google and Mirendil each get
The partnership gives Mirendil the computing foundation for its self-improving AI research. It also gives Google a strategic partner working on frontier recursive self-improving AI, a technology that Google can eventually shop around to enterprise customers.
Neyshabur said Mirendil’s software and systems layer can help customers get more out of Google’s hardware. If that works, Google gains more than a customer for its cloud infrastructure. It gains a company building technology that could make its AI hardware more useful to others.
For Mirendil, the deal supports a research direction that is closely tied to scientific discovery. The company believes self-improving AI could automate a lot of scientific and AI research, helping scientists make progress in medicine, biology and materials science.
Major labs such as Anthropic, where Mirendil’s co-founders hail from, have also worked on recursive self-improvement. Startups including Recursive Superintelligence and Ricursive Intelligence have recently emerged around the same goal. Mirendil’s Google Cloud deal shows how quickly the race is becoming an infrastructure race as much as a research race.
The larger AI infrastructure pattern
The agreement fits into a simple reality for the current AI market: companies pursuing advanced AI need dependable access to large amounts of compute. Without that access, even promising research ideas can be limited by training bottlenecks and hardware availability.
Mirendil’s deal also shows why cloud providers are competing for AI startups. A company building frontier AI can become a major infrastructure customer, a showcase for a cloud platform and a potential source of technology that enterprises may later want to use.
For now, the key fact is clear. Mirendil has secured a multi-year Google Cloud partnership worth upwards of $100 million, giving it access to TPUs, Nvidia GPUs and managed training clusters. That compute base will support its attempt to build AI that can keep doing research, improving its knowledge and advancing toward ambitious goals over time.