Mirendil's $100M Google Cloud Deal Powers AI Developer Tools Startups' Self-Improving AI
Mirendil just signed a multiyear deal with Google Cloud worth more than $100 million. The startup's goal? Build AI that gets better at its own job over time—forever. Behnam Neyshabur, Mirendil's co-founder and CEO, confirmed the number exclusively to TechCrunch on August 6, 2026. It's roughly half of what the company raised in seed funding at a $1 billion valuation back in late June. The math checks out. These folks are serious about compute.
The deal gives Mirendil access to both Google's TPUs and Nvidia GPUs, plus managed training clusters. They'll use all of it to work on what they call self-improving AI—also known as recursive self-improvement. That's the kind of AI that iteratively improves itself. The startup's ambition is pretty audacious: they want their system to eventually take on the work of an entire frontier AI lab.
How Self-Improving AI Actually Works
Self-improving AI isn't new. Major labs have been working on it for a while. Anthropic—where both of Mirendil's co-founders come from, has been exploring the space. So have a handful of startups that recently popped up around this goal, including Recursive Superintelligence and Ricursive Intelligence.
The concept is straightforward in theory, brutal in practice. You build an AI system that can look at its own performance, identify weaknesses, and then improve itself. Repeat. Over time, the system should get better at whatever domain you point it at.
Neyshabur put it bluntly: "You can have a self-improving AI where you can point a problem at it and it keeps getting better with time." He went on to ask the rhetorical question that really matters: "How can we have an AI system that keeps doing research, keeps improving its own knowledge and performance for Alzheimer's disease?" That's the kind of ambition that makes investors write checks.
The technology targets automation of scientific research across medicine, biology, and materials science. If it works, it could fundamentally change how we approach complex problems.
The Compute Problem, and Why Google Matters for AI Developer Tools Startups
Training self-improving AI requires enormous amounts of computing power. That's not controversial. What's interesting is how Mirendil plans to handle it.
Harsh Mehta, Mirendil's co-founder, said training is increasingly about matching the right workloads to the right hardware. "These models are really good at working with different workloads and chips, and assigning the right workloads to the right chips," he said. "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's the key. Google provides multiple kinds of chips, TPUs, Nvidia GPUs. Mirendil's software and systems layer helps customers get more out of that hardware. It's a software play on top of infrastructure. The startup optimizes how you use Google's hardware, and Google gets a strategic partner building frontier recursive self-improving AI that it can eventually shop around to enterprise customers.
Amin Vahdat, Google's SVP and chief technologist of AI and infrastructure, put it this way: AI advancement isn't just about chip-level performance anymore. It's about "how we orchestrate entire systems of intelligence and break through the physical constraints of scaling."
The Bigger Picture for AI Developer Tools Startups and India
This deal is part of a broader trend shaping the AI industry. Cloud giants are courting startups with huge infrastructure commitments. AI companies are snatching up as many compute deals as they can to secure access as they scale. Both trends are happening right now, and they're not slowing down.
Mirendil sits at the intersection of several important shifts. The company is building infrastructure for AI developer tools startups that want to push boundaries. The founding team brings deep expertise from Anthropic. The technical focus on recursive self-improvement could reshape how we think about scientific research. And the compute deal with Google, worth $100 million, roughly half a billion-dollar seed round, signals that the market is betting big on this approach.
For India's growing AI ecosystem, this could be a signal worth paying attention to. The country has been building out its tech services infrastructure for years. Companies like HCL are now entering the AI datacenter business. A startup like Mirendil, with its focus on maximizing compute efficiency across heterogeneous hardware, could find partners there. The question is whether the broader ecosystem will follow.
The deal itself is multiyear, which means Mirendil's not looking for quick wins. They're building something that could fundamentally change how frontier AI research gets done. Whether that's good or bad depends on who you ask. The compute bills alone suggest they're not messing around.
What This Means for the Market
A $100 million compute deal isn't cheap. But when you're building self-improving AI, one of the most compute-hungry endeavors in existence, it's almost necessary. The startup's seed round at $1 billion valuation tells you what the market thinks about the opportunity.
Mirendil's approach is interesting because it's not just about building another model. It's about building a system that builds better systems. That's recursive self-improvement, and it's the holy grail for anyone who believes AI will eventually automate scientific discovery.
The partnership with Google gives Mirendil the infrastructure it needs. Whether it delivers on the ambition remains to be seen. But at least one thing is clear: the companies building tools for AI developers, the startups raising venture capital, and the cloud giants competing for compute contracts are all converging on the same problem. How do you scale intelligence when the hardware keeps hitting physical limits?
Mirendil's answer involves self-improving AI, heterogeneous compute, and a hell of a lot of money. Time will tell if it works.
Related coverage: