Google notes many large AI labs already use its infrastructure, and Mirendil’s deployment expands Google’s customer base in a market competing for access to both proprietary and third-party chips.

At the core is recursive self-improvement, aiming to automate and accelerate scientific research in medicine, biology, and materials science by enabling AI systems to continually enhance their knowledge and capabilities.

The arrangement includes managed training clusters to support Mirendil’s recursive self-improvement AI work and its end-to-end training workflows.

The collaboration positions Google to gain a strategic partner in building frontier recursive self-improving AI, with potential future commercialization to enterprise customers.

The deal reflects a broader industry trend of AI labs avoiding dependency on a single chip architecture to optimize performance, software compatibility, and resources.

Mirendil has chosen Google Cloud’s AI Hypercomputer, gaining access to both Google TPUs and Nvidia-based infrastructure to advance AI development and research.

This access enables Mirendil to match workloads with the most suitable computing architecture, combining TPUs and NVIDIA GPUs as tasks require.

Behnam Neyshabur, Mirendil’s co-founder and CEO, emphasizes that deploying with Google Cloud will accelerate the research loop and broaden access for scientists and engineers.

The deal’s valuation signals substantial infrastructure investment for AI startups, roughly half of Mirendil’s seed funding at a billion-dollar valuation earlier in the summer.

The deployment provides flexible computing resources to accelerate parts of the research cycle and make frontier AI capabilities more accessible to a wider scientific community.

Mirendil’s software and systems layer is expected to optimize hardware utilization, potentially giving Google a competitive edge in frontier AI partnerships beyond hardware alone.

Google frames its infrastructure as orchestrating whole systems, not just raw chip performance, to overcome scaling constraints in frontier AI.