AI research firm Mirendil has formed a multi-year alliance with Google Cloud to obtain substantial computing resources for its work on self-improving AI, as exclusively reported by TechCrunch.

The agreement reflects key patterns in the AI sector where major cloud providers offer large-scale infrastructure support to emerging companies while AI firms secure extensive hardware access to support rapid growth.

Valued at more than $100 million, the partnership was confirmed by Mirendil co-founder and CEO Behnam Neyshabur. This amount represents about half of the seed funding the company secured at a $1 billion valuation in late June.

The arrangement provides Mirendil with access to Google TPUs, Nvidia GPUs, and managed training clusters to advance its self-improving AI initiatives. The company aims for its systems to eventually handle the full scope of work performed by leading AI research labs.

Self-improving AI, or recursive self-improvement, involves systems that refine their own capabilities over time. This idea has been explored by major organizations including Anthropic, the prior employer of Mirendil's co-founders, and newer entrants such as Recursive Superintelligence and Ricursive Intelligence.

Mirendil expects this approach to automate significant portions of scientific and AI research, supporting advances in areas including medicine, biology, and materials science.

Neyshabur described how AI could replicate the way human researchers acquire knowledge across new fields, build expertise, and enhance their results over time. He noted that a self-improving system could be directed toward a challenge and continue to progress independently.

He added that the technology could enable AI to conduct ongoing research on complex issues such as Alzheimer's disease, setting ambitious targets and sustaining improvement without constant human intervention.

Co-founder Harsh Mehta emphasized that training these models demands vast computing power and increasingly requires efficient matching of tasks to appropriate hardware. He highlighted how the models can handle varied workloads across different chips, allowing Google to supply multiple accelerator types that reduce costs for both the lab and its end users.

This adaptability forms a core element of Google's infrastructure strategy. Amin Vahdat, SVP and chief technologist of AI and infrastructure at Google, stated that progress in AI now depends on coordinating full systems of intelligence rather than isolated chip performance alone, in order to overcome physical scaling limits.

Neyshabur explained that Mirendil's software and systems optimize the use of Google's hardware, offering the cloud provider an additional advantage against competitors. In exchange, Google gains a partner focused on frontier recursive self-improving AI that could later be offered to enterprise clients.