Mirendil Lands $100M+ Google Cloud Deal to Power Self-Improving AI for Scientific Discovery

A frontier AI startup founded by former Anthropic researchers just locked in one of the largest cloud infrastructure deals of the year. Mirendil has signed a multiyear partnership with Google Cloud valued at upward of $100 million to secure the massive compute capacity it needs to develop self-improving AI systems designed to accelerate scientific discovery .

The deal, reported exclusively by TechCrunch on August 6, 2026, is roughly half of the $200 million seed round the startup raised two months earlier at a $1 billion valuation . It gives Mirendil access to Google Cloud's AI Hypercomputer, which includes both Google's custom Tensor Processing Units (TPUs) and Nvidia GPUs, along with managed training clusters .

Here is the full breakdown of the partnership, what it means for the startup, and why it matters for the broader AI and cloud computing landscape.

The Founders and Their Mission

Mirendil was co-founded by Behnam Neyshabur (CEO) and Harsh Mehta, both former researchers at Anthropic who previously worked at Google . They left Anthropic to start Mirendil with a bold mission: build AI systems that can accelerate their own improvement and drive scientific discovery.

The startup's goal is to automate scientific research — creating self-improving AI designed to accelerate discovery in fields like medicine, biology, chemistry, and materials science . The core idea is that AI models can not only assist with experiments but also generate hypotheses, design protocols, and iteratively improve their own capabilities, effectively speeding up the entire research cycle.

The Deal: Financial Terms and Compute Infrastructure

The partnership is structured as a multiyear commitment valued at more than $100 million . Here are the key details:

• Value: Upward of $100 million, approximately half of Mirendil's $200 million seed round .
• Compute Resources: Access to Google Cloud's AI Hypercomputer, which includes both Google's own TPUs and Nvidia GPUs, plus managed training clusters .
• Timing: Announced two months after Mirendil's seed round closed at a $1 billion valuation in late June 2026 .

This scale of commitment reflects the enormous compute demands of frontier AI research. For context, the $100M+ figure makes it one of the largest cloud partnerships for an AI startup this year .

Technical Approach: Matching Workloads to Hardware

To keep costs manageable, Mirendil employs a flexible compute strategy — it matches different AI training workloads to the most cost-effective hardware . The startup uses:

• Google's TPUs for certain training tasks
• Nvidia GPUs for others

This workload-matching approach reduces waste and lets the startup get more research done within its $100M+ commitment. It's a practical optimization that matters for any AI lab trying to stretch its compute budget.

How Both Sides Benefit

The partnership creates clear value for both Mirendil and Google Cloud.

For Mirendil and its customers: The deal guarantees the massive, reliable compute capacity needed to train and run self-improving models. This allows Mirendil to offer its customers faster scientific research cycles — without those customers having to manage their own infrastructure. The startup's AI is designed to eventually tackle problems in pharmaceutical development, materials discovery, and other research-intensive industries.

For Google Cloud: Google secures a high-profile frontier AI partner that anchors a major enterprise cloud deal . It helps Google compete with AWS and Azure for AI-native startups by demonstrating that Google Cloud's TPU + GPU stack can support cutting-edge research. It also gives Google early insight into the infrastructure needs of self-improving AI, which can inform future hardware and cloud product roadmaps .

As one analysis noted, the deal signals Google's growing appetite for backing cutting-edge AI research beyond its own labs .

What This Means for the AI Industry

Mirendil's deal is part of a broader trend: cloud giants are increasingly courting AI startups with large infrastructure commitments, locking in long-term relationships before competitors can step in. The $100M+ figure matches the scale of similar deals from AWS and Azure with other frontier labs.

For the self-improving AI space, this partnership provides concrete infrastructure backing for a research area that remains largely theoretical. As the TechCrunch article noted, Mirendil's AI is not yet proven at scale, but the compute capacity now exists for it to try .

The deal also highlights how much compute frontier AI research requires — and how much cloud providers are willing to invest to win those workloads.