Mirendil inks $100M Google Cloud deal for self-improving AI
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Mirendil signed a multiyear Google Cloud deal worth upward of $100 million to source compute for its self-improving AI research, per a TechCrunch exclusive.
- The commitment represents roughly half of Mirendil's seed funding, which was raised at a $1 billion valuation in late June, and grants the startup access to Google TPUs, Nvidia GPUs, and managed training clusters.
- Mirendil is pursuing recursive self-improvement — AI that iteratively improves itself — with co-founders Behnam Neyshabur and Harsh Mehta saying it could automate scientific and AI research in medicine, biology, and materials science.
- Both co-founders previously worked at Anthropic; other startups chasing the same goal include Recursive Superintelligence and Ricursive Intelligence.
- Google's Amin Vahdat framed the partnership as orchestration of "entire systems of intelligence" rather than chip-level performance alone, while Neyshabur said Mirendil's software stack helps customers extract more from Google's hardware.
- The strategic angle for Google is a frontier-AI partner building recursive self-improvement technology it can eventually resell to enterprise customers, per Neyshabur.
Why it matters: The $100M-plus commitment consumes roughly half of Mirendil's seed round in compute alone, underscoring how capital-intensive recursive self-improvement research has become before the startup has shipped a product. Google locks in a frontier-AI partner whose technology it can resell to enterprises, while Mirendil locks in access to the TPUs and Nvidia GPUs needed to train compute-hungry self-improving models.
