Mirendil Secures $100 Million Google Cloud Deal for Self-Improving AI

Mirendil, a San Francisco-based AI research lab, has entered a multi-year partnership with Google Cloud valued at more than $100 million. The agreement provides the startup with the massive compute necessary to develop AI systems capable of recursive self-improvement.

To power its operations, Mirendil is deploying Google Cloud's AI Hypercomputer infrastructure. This hybrid setup combines Nvidia GPUs with Google's proprietary Tensor Processing Units (TPUs), specifically the v5P chips. This dual-hardware approach allows the lab to match specific pre-training and post-training workloads to the most efficient hardware available, optimizing both performance and cost. The lab was founded by CEO Behnam Neyshabur and CTO Harsh Mehta, both of whom previously served as researchers at Anthropic.

The Push for Recursive Improvement

Mirendil is positioning itself as a catalyst for scientific discovery rather than a provider of consumer chatbots. The lab focuses on "recursive self-improvement," aiming to create systems that can independently conduct and accelerate research in complex fields such as materials science, medicine, and biology. By automating the cycle of hypothesis, experiment, and evaluation, the company hopes to democratize frontier R&D, making high-level research tools accessible to specialized institutes and universities.

Behnam Neyshabur has emphasized that human speed has historically been the primary bottleneck in AI development. "Progress in AI has been bounded by how fast humans can run the research loop," Neyshabur stated, adding that the company is building systems to accelerate and improve that loop itself. He further noted that a self-improving AI allows a user to "point a problem at it and it keeps getting better with time."

Industry Implications

This partnership signals a strategic pivot toward "AI for AI research." While the previous wave of generative AI focused on human-facing interfaces, Mirendil's approach treats the AI as the primary researcher. The scale of the deal also underscores the critical necessity of flexible, multi-chip infrastructure. As models grow in complexity, the ability to toggle between TPUs and GPUs is becoming a prerequisite for labs attempting to maintain a competitive edge without incurring unsustainable compute costs.

Scaling the Frontier

The infrastructure deal follows a massive capital injection for the startup. Mirendil launched with a $200 million seed round at a $1 billion valuation, a funding effort co-led by Kleiner Perkins and Andreessen Horowitz, with participation from Nvidia.

As Mirendil integrates the AI Hypercomputer into its workflow, the industry will be watching to see if the lab can successfully transition from theoretical recursive improvement to tangible breakthroughs in the physical sciences. The primary metric for success will be whether these self-improving systems can produce verified scientific discoveries faster than traditional human-led research teams.