The race to build AI that can improve itself just got a significant infrastructure boost. Mirendil, a frontier AI lab founded by former Anthropic researchers, has signed a multi-year partnership with Google Cloud worth upwards of $100 million. The deal gives the startup access to Google’s Tensor Processing Units (TPUs), NVIDIA GPUs, and managed training clusters to develop what it calls “self-improving AI”.
Announced on August 6, 2026, the agreement represents roughly half of the $200 million seed funding Mirendil raised at a $1 billion valuation in late June. For a company that officially launched just months ago, securing this level of compute commitment signals both the capital intensity of modern AI development and the urgency with which startups are racing to secure infrastructure.
But beneath the headline numbers lies a more interesting story about how the AI industry is evolving—and what happens when the people who helped build frontier models decide to build something that could eventually replace them.
What Mirendil Is Actually Building
Mirendil’s core ambition is recursive self-improvement: AI systems that iteratively enhance their own capabilities without requiring constant human intervention. Co-founder and CEO Behnam Neyshabur describes it as AI that can “keep getting better with time” when pointed at a problem. The startup hopes its technology will eventually handle the work of an entire frontier AI lab.
The practical applications span scientific research—medicine, biology, and materials science are specific targets. Neyshabur frames the vision around ambitious 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?”
This isn’t just theoretical. The founders—Neyshabur and Harsh Mehta—met at Google in 2019, moved to Anthropic in late 2024, and left in December 2025 shortly after Claude Opus 4.5 launched. Their founding team includes 20 researchers from Anthropic, xAI, Google DeepMind, and OpenAI. They’re building on concepts that major labs have been exploring internally.
Why This Deal Matters for Google Cloud
For Google, the partnership is strategic beyond the immediate revenue. Amin Vahdat, Google’s SVP and chief technologist of AI and infrastructure, framed AI advancement as no longer just about “chip-level performance” but “how we orchestrate entire systems of intelligence and break through the physical constraints of scaling”.
The deal gives Mirendil access to Google’s AI Hypercomputer—a mix of TPU accelerators and full-stack NVIDIA infrastructure. The startup is already live with a cluster of TPU v5P chips, with NVIDIA systems coming online soon.
But there’s a reciprocal benefit that makes this arrangement particularly interesting. Neyshabur noted that Mirendil’s software and systems layer helps customers extract more value from Google’s hardware. In effect, Google gets a strategic partner building frontier recursive AI that could eventually be packaged and sold to enterprise customers. Mirendil becomes both a customer and a potential distribution channel for Google’s infrastructure.
The Compute Arms Race Intensifies
This deal mirrors two trends reshaping the AI industry. Cloud providers are aggressively courting startups with massive infrastructure commitments, often structuring deals that lock in spending over multiple years. Meanwhile, AI companies are securing as many compute agreements as they can, treating compute access as a strategic resource nearly as valuable as talent or data.
Mirendil isn’t alone in this pursuit. Recursive Superintelligence, another self-improving AI startup, recently signed a $400 million compute deal with Amazon. These aren’t isolated transactions—they’re part of a broader pattern where cloud giants are placing bets on specific approaches to AI development, hoping to capture the infrastructure spending that will follow if any of these bets pay off.
For Mirendil, the timing matters. The company raised $200 million in seed funding—one of the largest seed rounds in AI history. Committing roughly half of that to compute infrastructure through a single cloud provider is a significant concentration of resources. It suggests either deep conviction in Google’s hardware roadmap or a recognition that securing capacity now is more important than maintaining flexibility across multiple providers.
A Contested Technical Frontier
Recursive self-improvement remains a deeply contested area of AI research. Anthropic, where Mirendil’s founders previously worked, has identified it as a potential danger—the theory being that a model rewriting its own code without oversight could slip beyond human control.
Mirendil’s founders see it differently. They frame recursive self-improvement as the “shortest path” to faster science—a problem that can be supervised rather than avoided. This philosophical divide matters because it shapes how each organization approaches safety, transparency, and the pace of deployment.
The tension is also commercial. Anthropic’s terms of service reportedly forbid using its tools to build competing services, a policy the company has defended as standard among model providers. Mirendil is explicitly trying to build what the big labs keep for themselves. As Andreessen Horowitz partner Matt Bornstein put it, the labs are being “rational economic actors” when they deny customers the means to supercharge their own models.
What This Means for Scientific Research
The most immediate implication of Mirendil’s work is for scientific discovery. The startup’s pitch is straightforward: automate the research loop so scientists can focus on questions rather than implementation. A university biology lab could use Mirendil’s platform to build a drug-target model without needing a dedicated machine-learning team. Work that currently takes months could compress into days.
This is the kind of productivity leap that could reshape how research is conducted across multiple disciplines. But it also raises questions about who controls the means of AI research. If self-improving AI becomes a commodity that any institution can access, the barrier to entry for cutting-edge research drops dramatically. If it remains concentrated among a few well-funded players, the gap between haves and have-nots widens.
Mirendil’s model appears aimed at the former outcome—democratizing access to frontier AI research capabilities. Whether that vision survives the economics of scaling remains to be seen.
Infrastructure as Strategy
Perhaps the most important aspect of this deal is what it reveals about how the AI industry is structuring itself. Cloud providers aren’t just selling compute—they’re buying into specific technical visions. Google’s partnership with Mirendil represents a bet that recursive self-improvement is a viable path forward, and that being the infrastructure provider for that path has long-term value.
For Mirendil, the deal provides something money alone can’t always buy: priority access to cutting-edge hardware at a time when everyone is competing for the same scarce resources. The managed training clusters and end-to-end infrastructure partnership Google is providing suggest this goes beyond a standard customer agreement.
The question is whether this concentration of compute into a single cloud provider creates dependency or accelerates development. Mirendil’s founders clearly believe the flexibility of Google’s multi-chip approach—matching workloads to the right accelerators—outweighs the risk of lock-in. The next 12 to 18 months will test that thesis.
Mirendil’s $100 million-plus deal with Google Cloud is more than a large infrastructure commitment. It’s a signal that recursive self-improvement—once confined to theoretical discussions and internal lab research—is becoming a commercially funded, infrastructure-backed pursuit.
The startup’s founders left Anthropic to build what they believe is the next frontier of AI: systems that can accelerate their own development. Google has placed a significant bet on that vision, providing the compute capacity needed to test whether self-improving AI can move from concept to reality.
For scientists in fields like medicine and materials science, the potential payoff is enormous. For the AI industry, the implications are even larger. If Mirendil succeeds, the distinction between AI labs and AI itself may begin to blur—and the companies that control the infrastructure for that transition will shape what comes next.
