Mirendil Google Cloud deal reportedly tops $100 million for AI compute

Mirendil’s Google Cloud deal is reportedly worth more than $100 million and gives the AI lab a multi-year supply of computing infrastructure for its self-improving-AI research. For investors watching the AI buildout, the arrangement is another example of a young lab making a large commitment to secure the capacity it says it needs for training.

TechCrunch reported the agreement on August 6, citing Mirendil co-founder and CEO Benham Neyshabur. Neither a Mirendil announcement nor a Google Cloud announcement confirming the reported arrangement was included in the available materials.

According to TechCrunch, Mirendil will have access to Google’s tensor processing units, or TPUs, Nvidia graphics processing units, or GPUs, and managed training clusters. Co-founder Harsh Mehta told the publication that the flexibility to assign different workloads to different types of hardware could help the company manage costs.

What does the Mirendil Google Cloud deal include?

The reported partnership covers three categories of infrastructure: Google TPUs, Nvidia GPUs and managed training clusters. TechCrunch described it as a compute-capacity agreement intended to support Mirendil’s work on self-improving AI.

Terms beyond the reported value, duration and infrastructure access were not disclosed in the material available. The value was described by Neyshabur as upward of $100 million, rather than as a precisely stated contract total.

What is self-improving AI?

Self-improving AI, also called recursive self-improvement in TechCrunch’s report, refers to systems designed to improve themselves through repeated iterations. Mirendil’s stated aim is to develop systems that can keep building knowledge and improve their performance on a given problem over time.

The company believes the approach could automate parts of scientific and AI research, including work connected with medicine, biology and materials science. That is a research goal, not a demonstrated result. Mirendil also hopes its systems could eventually handle work now performed by an entire frontier AI lab, TechCrunch reported.

TechCrunch reported that training self-improving AI requires enormous computing power. Mirendil’s move toward a mix of Google and Nvidia hardware reflects its view that different workloads can be matched with different accelerators.

How large is the commitment relative to Mirendil’s funding?

TechCrunch said the reported deal value is about half of Mirendil’s late-June seed financing at a $1 billion valuation. Separately, a LinkedIn search snippet attributed to Neyshabur described Mirendil as announcing a $200 million seed round to build a platform for scientists to create their own AI.

The reported cloud commitment does not establish that Mirendil has achieved its research objectives. It does show the scale of infrastructure the company says it is lining up to pursue them.

This story draws on original reporting from TechCrunch.