AI lab Mirendil has signed a multiyear deal with Google Cloud worth more than $100 million to fuel its work on self-improving AI, according to TechCrunch AI, which reported the partnership as an exclusive. Co-founder and CEO Behnam Neyshabur confirmed the figure to TechCrunch AI. That’s roughly half of what the startup raised in its seed round at a $1 billion valuation back in late June, which tells you how much compute now costs at the frontier.
Here’s what the deal actually buys. Mirendil gets access to Google’s TPUs, Nvidia GPUs, and managed training clusters. In return, Google gets a strategic partner building recursive self-improving AI, technology it can eventually sell to enterprise customers.
What Mirendil is building
Self-improving AI, also called recursive self-improvement, means systems that iteratively make themselves better. Point a hard problem at the model, and it keeps improving its own knowledge and performance over time. Mirendil’s co-founders came out of Anthropic, where the concept has been an active research thread, and the startup’s ambition is blunt: it wants an AI that can eventually do the work of an entire frontier AI lab.
Neyshabur frames it around science. He thinks AI can mimic how human researchers learn a new field, build up expertise, and gradually sharpen their results.
“You can have a self-improving AI where you can point a problem at it and it keeps getting better with time,” he told TechCrunch AI. His example was Alzheimer’s disease: an AI that “keeps doing research, keeps improving its own knowledge and performance” until it makes progress. The pitch extends to medicine, biology, and materials science.
Why this matters
This deal sits on top of two trends reshaping the industry, as TechCrunch AI notes:
- Cloud giants are courting startups with massive infrastructure commitments to lock in future customers.
- AI companies are grabbing every compute deal they can to secure access before they scale.
What stands out here is the money math. A company spends half its seed round on compute before it has a product in market. That’s the new cost of entry for anyone serious about training frontier models, and it explains why cloud partnerships now read like the real fundraising events.
Mirendil isn’t alone in chasing recursive self-improvement. Startups like Recursive Superintelligence and Ricursive Intelligence have popped up around the same goal, and the big labs are already deep in it. The idea has moved from research paper to funded race.
The hardware angle
Co-founder Harsh Mehta says training is increasingly about matching the right workloads to the right chips. Google offers several kinds of accelerators, and Mehta argues that flexibility lets Mirendil mix and match jobs to hardware and cut costs, both for itself and for its own customers.
“These models are really good at working with different workloads and chips, and assigning the right workloads to the right chips,” Mehta told TechCrunch AI.
That flexibility is central to Google’s pitch. Amin Vahdat, SVP and chief technologist of AI and infrastructure at Google, said in a statement that progress isn’t just about chip-level speed anymore, “but how we orchestrate entire systems of intelligence and break through the physical constraints of scaling.”
There’s a two-way benefit worth naming. Neyshabur says Mirendil’s software layer helps customers squeeze more out of Google’s hardware, which hands the cloud giant another edge against AWS and Microsoft. Google gets better utilization and a frontier showcase. Mirendil gets the chips it can’t afford to buy outright.
What to watch next
The promises here are big, and the technology is early. Recursive self-improvement remains largely a research bet, not a shipping product, so treat the Alzheimer’s-scale claims as direction rather than delivery. The signal that matters is the spending pattern: expect more nine-figure compute deals between cloud providers and well-funded labs, and expect more of a startup’s raise to disappear into GPU and TPU time before anything reaches customers.
If you’re building in AI, the takeaway is practical. Compute access is becoming the moat, and the companies locking in multiyear capacity now are the ones betting they’ll still be standing when the self-improving systems actually work. Full details are in the original TechCrunch AI report.