Mirendil, an AI lab focused on self-improving AI, has signed a multi-year partnership with Google Cloud worth upwards of $100 million. The deal secures compute capacity for the startup's research into recursive self-improvement, a field where AI systems aim to enhance their own capabilities.
The agreement gives Mirendil access to Google's TPUs and Nvidia GPUs, along with managed training clusters. The startup's co-founders, who previously worked at Anthropic, believe this infrastructure will help them pursue an ambitious goal: building AI that can eventually take on the work of an entire frontier AI lab.
A Big Bet on Compute
Training self-improving AI requires enormous computing power, and Mirendil's deal reflects that reality. The $100 million-plus commitment is roughly half of the startup's seed funding, which it raised at a $1 billion valuation in late June. That seed round valued the company at $1 billion, a striking figure for a lab that has yet to ship a commercial product.
Benham Neyshabur, co-founder and CEO of Mirendil, confirmed the deal's value and outlined the company's vision. "You can have a self-improving AI where you can point a problem at it and it keeps getting better with time," he said. Neyshabur thinks AI can mimic how human scientists learn and improve, pointing to a concrete example. "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 asked.
The CEO believes self-improving AI can be pointed at a problem and keep improving indefinitely. "This technology allows us to set goals that are ambitious for AI, and the AI would keep making progress," he said.
Hardware Flexibility at the Core
Mirendil's software and systems layer helps customers get more out of Google's hardware, according to Neyshabur. That layer is central to how the startup plans to use its new compute capacity.
Harsh Mehta, co-founder of Mirendil, discussed how training workloads are shifting. "These models are really good at working with different workloads and chips, and assigning the right workloads to the right chips," he said. Mehta claims Google's multiple chip types allow for mixing and matching workloads to lower costs. "[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," he added.
That flexibility is central to Google's broader pitch to AI companies. Amin Vahdat, SVP and chief technologist of AI and infrastructure at Google, framed the company's approach in terms of orchestration rather than raw chip performance. "But how we orchestrate entire systems of intelligence and break through the physical constraints of scaling," he said, describing what he sees as the key to AI advancement.
Stay ahead of the AI curve
The most important updates, news, and content — delivered weekly.
No spam. Unsubscribe anytime.
A Crowded Field Emerges
Mirendil is not alone in chasing recursive self-improvement. Anthropic, where Mirendil's co-founders hail from, has been working on self-improving AI. Startups like Recursive Superintelligence and Ricursive Intelligence have recently emerged in this space, each focused on similar goals.
Mirendil believes self-improving AI will automate scientific and AI research, with fields like medicine, biology, and materials science potentially benefiting. The startup's pitch is that an AI system could continuously push forward on hard problems without human intervention.
The deal mirrors two broader trends: cloud giants courting startups with huge infrastructure commitments, and AI companies securing compute deals to scale. Google gets a strategic partner building frontier recursive self-improving AI, which it can eventually shop to enterprise customers. Mirendil's software layer gives Google a potential leg up in the race against competition.
What Comes Next
Mirendil raised seed funding at a $1 billion valuation in late June, and now it has locked in compute capacity for years to come. The startup's co-founders are betting that their software layer, combined with Google's hardware, will let them train models that improve themselves.
Neyshabur's vision is clear: an AI that keeps getting better with time, pointed at problems like Alzheimer's disease, and eventually capable of running an entire frontier AI lab. Whether that vision materializes depends on whether the compute, the software, and the research all come together.
For now, Mirendil has secured the infrastructure it says it needs. The $100 million-plus deal is a statement of intent, backed by a $1 billion valuation and a team with roots at one of the world's leading AI labs.