AI lab Mirendil has entered into a multi-year partnership with Google Cloud to source computing power for self-improving AI research, TechCrunch has learned exclusively.
The deal reflects two trends shaping the AI industry. Cloud giants are courting startups with huge infrastructure commitments. Second, AI companies are acquiring as many compute deals as possible to ensure access as they scale.
Mirendil co-founder and CEO Behnam Neyshabur told TechCrunch that the deal is worth more than $100 million. This is about half the amount Mirendil raised in seed funding in late June at a valuation of $1 billion.
The deal gives the startup access to both Google’s TPUs and Nvidia GPUs, as well as a managed training cluster used by Mirendil as it works on self-improving AI. The startup hopes its AI will eventually be able to take over the work of the entire Frontier AI Lab.
Self-improving AI, also known as recursive self-improvement, refers to AI systems that iteratively improve themselves. It’s a concept that major labs like Anthropic, where Mirendil’s co-founders are from, have been working on it. A handful of startups, including Recursive Superintelligence and Ricursive Intelligence, have also recently sprung up to help achieve that goal.
Mirendil believes this process will automate much scientific and AI research, helping scientists advance in fields such as medicine, biology and materials science.
Neishavar believes that AI can mimic the way human scientists learn more about new areas, accumulate knowledge and expertise, and gradually improve their performance. “With self-improving AI, you can point out problems and continue to improve over time,” he said.
“How can we have an AI system that continues to research and improve its own knowledge and performance when it comes to Alzheimer’s disease?” he continued. “This technology will allow us to set ambitious goals for AI, and it will continue to advance.”
However, training self-improving AI requires huge amounts of computing power. Harsh Mehta, co-founder of the institute, said training is increasingly about matching the right workload to the right hardware.
“These models are very good at handling different workloads and chips and assigning the right workload to the right chip,” Mehta says. “[Google]offers multiple types of chips. This flexibility ultimately allows us to match workloads with the right type of accelerator, lowering costs not only for us but also for customers using our systems.”
This flexibility is at the heart of Google’s AI infrastructure proposition. Amin Vahdat, Google’s SVP and chief technologist for AI and infrastructure, said in a statement that advances in AI are no longer just about chip-level performance, but “how we orchestrate the entire intelligence system and break through the physical constraints of scaling.”
Neishaboul said Mirendil’s software and system layers will help customers get the most out of Google’s hardware, potentially giving the cloud giant an additional advantage over its rivals. In return, Google acquired a strategic partner to build cutting-edge recursive self-improvement AI technology that it could eventually sell to enterprise customers.
If you buy through links in our articles, we may earn a small commission. This does not affect editorial independence.
