Exclusive: Mirendil inks $100M+ Google Cloud deal to scale self-improving AI

In a significant move that underscores the intensifying arms race for computational resources, AI startup Mirendil has secured a multi-year partnership with Google Cloud. The deal, valued at upwards of $100 million, is designed to provide the lab with the massive compute capacity required to advance its research into recursive, self-improving artificial intelligence.

This agreement highlights two dominant narratives currently defining the AI landscape: the aggressive pursuit of infrastructure by well-funded startups and the strategic efforts of cloud giants to lock in partnerships with the next generation of frontier AI developers.

Fueling Recursive Intelligence

For Mirendil, the capital commitment represents a substantial portion of its recent $1 billion valuation seed round, which closed in late June. According to co-founder and CEO Behnam Neyshabur, the partnership grants the startup access to a diverse array of hardware, including Google’s proprietary TPUs and Nvidia GPUs, alongside managed training clusters.

The core objective is to develop AI systems capable of iterative self-improvement—a concept that aims to automate the heavy lifting of scientific and AI research. By mimicking the way human experts accumulate knowledge and refine their methodologies, Mirendil hopes to create systems that can autonomously tackle complex challenges in fields like biology, medicine, and materials science.

"You can have a self-improving AI where you can point a problem at it and it keeps getting better with time," said Neyshabur. "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 technology allows us to set goals that are ambitious for AI, and the AI would keep making progress."

Strategic Hardware Orchestration

Training models that possess the ability to improve themselves requires an unprecedented scale of compute. Mirendil co-founder Harsh Mehta emphasized that the partnership is as much about efficiency as it is about raw power. The lab is focusing on optimizing how specific workloads are paired with the most suitable hardware.

• Hardware Flexibility: The ability to mix and match workloads across Google’s diverse chip offerings is a key advantage.
• Cost Efficiency: By assigning the right tasks to the right accelerators, Mirendil aims to lower operational costs for both its internal research and its future enterprise customers.
• Systems Layer: Mirendil’s proprietary software layer is designed to extract maximum performance from Google’s infrastructure, providing a competitive edge for both the startup and the cloud provider.

A Symbiotic Partnership

For Google, the deal is a strategic win. By supporting a lab focused on frontier recursive AI, Google is positioning itself as the primary infrastructure provider for the next wave of autonomous research tools. Amin Vahdat, SVP and chief technologist of AI and infrastructure at Google, noted that the industry has moved beyond simple chip-level performance metrics.

"It’s about how we orchestrate entire systems of intelligence and break through the physical constraints of scaling," Vahdat stated.

As Mirendil continues to build out its systems, the startup joins a growing cohort of labs—including Anthropic (where Mirendil’s founders previously worked) and startups like Recursive Superintelligence—that are betting on the future of self-improving models. With this $100 million infusion of compute power, Mirendil is now firmly equipped to test whether its AI can eventually perform the work of an entire frontier research laboratory.