Mirendil commits more than $100M to Google Cloud AI compute
AI lab Mirendil has signed a multi-year partnership with Google Cloud worth more than $100 million to secure computing capacity for self-improving AI research. Co-founder and CEO Benham Neyshabur told TechCrunch that the agreement provides access to Google TPUs, Nvidia GPUs and managed training clusters.
The commitment is roughly half the seed funding Mirendil raised in late June at a $1 billion valuation. It gives the startup infrastructure for work on systems intended to iteratively improve their own knowledge and performance.
Compute flexibility for recursive self-improvement
Mirendil describes self-improving AI, also called recursive self-improvement, as AI that can repeatedly improve itself. The company ultimately wants its technology to take on the work of an entire frontier AI lab.
Neyshabur said the intended model is an AI system that can be assigned a problem and become better over time. He cited Alzheimer’s disease as an example of a domain in which a system could continue research and improve its knowledge and performance.
Co-founder Harsh Mehta said the training challenge increasingly involves placing the right workloads on the right hardware. Google Cloud’s mix of chip types lets Mirendil match workloads with suitable accelerators, a flexibility Mehta said could lower costs for both the lab and customers using its systems.
Why the infrastructure agreement matters
The arrangement reflects a wider contest for AI infrastructure: cloud providers are making substantial commitments to startups, while AI companies seek multiple sources of capacity as their requirements grow. Mirendil’s founders previously worked at Anthropic, where related research has also been pursued.
For enterprises assessing AI-agent strategies, enterprise AI agents without frontier labs shows how enterprise AI agents are being positioned without reliance on frontier labs, while Mirendil’s agreement highlights the separate importance of durable infrastructure access for research-intensive systems.
Amin Vahdat, Google’s SVP and chief technologist of AI and infrastructure, said progress depends not only on chip-level performance but also on orchestrating entire systems of intelligence and overcoming the physical constraints of scale. Mirendil says its software and systems layer can help customers extract more from Google hardware.
The practical business implication is that organisations developing or buying advanced AI systems should evaluate accelerator choice, workload orchestration and long-term compute availability together, because each affects the ability to train and operate demanding models.