Mirendil inks $100M Google Cloud deal for self-improving AI

AI research lab Mirendil has entered into a multiyear partnership with Google Cloud valued at more than $100 million. The agreement secures critical computing power for the startup as it develops recursive self-improving AI systems. This massive infrastructure commitment represents roughly half of the capital Mirendil secured during its seed funding round in late June, which valued the company at $1 billion.

Under the terms of the agreement, Mirendil gains access to Google's proprietary Tensor Processing Units (TPUs) alongside Nvidia GPUs, managed training clusters, and specialized software tools. Co-founded by former Anthropic researchers Behnam Neyshabur and Harsh Mehta, the startup plans to use this mixed-hardware environment to build models that can iteratively upgrade their own capabilities. Mehta noted that the flexibility of using multiple chip architectures allows the startup to match specific workloads to the most efficient accelerators, ultimately driving down operational costs for both Mirendil and its future enterprise clients.

For AI practitioners and scientists, Mirendil's pursuit of self-improving AI aims to automate complex research workflows. The startup envisions systems that can independently accumulate domain expertise, learn from new data, and systematically solve difficult problems in fields like biology, materials science, and medicine. Neyshabur suggested that instead of relying solely on human engineers to manually tune and upgrade models, practitioners could deploy an AI system that 'keeps getting better with time' to tackle ambitious goals such as researching Alzheimer's disease.

The deal highlights an ongoing trend of cloud providers offering massive infrastructure pipelines to promising AI startups. By hosting Mirendil's workloads, Google Cloud positions itself to offer advanced recursive AI capabilities to its own enterprise clients in the future. Meanwhile, Mirendil's software layer will help optimize Google's underlying hardware, demonstrating how system-level orchestration is becoming just as vital as raw chip performance in the race to scale frontier AI.