Mirendil, an AI lab founded by former Anthropic and Google researchers, has signed a multiyear deal with Google Cloud worth more than $100 million. The contract, revealed on 6 August 2026 by TechCrunch, gives the startup access to TPUs and Nvidia GPUs to train self-improving AI systems.
Key takeaways
- Google Cloud deal: over $100 million, multiyear
- Scope: access to TPUs, Nvidia GPUs and managed training clusters
- The amount is roughly half of Mirendil's seed round (about $200M at a $1B valuation)
- Goal: self-improving AI (recursive self-improvement) to automate research
- Founders: Behnam Neyshabur (CEO) and Harsh Mehta, previously at Anthropic and Google
What the deal covers
Mirendil gains access to three parts of Google Cloud's infrastructure: Google's own TPUs, Nvidia GPUs, and managed training clusters that simplify running large training jobs. According to TechCrunch, the contract is worth roughly half of the seed round the startup closed in late June — about $200 million at a $1 billion valuation. That is an unusual scale of compute spending for such a young company.
What self-improving AI means
Mirendil is building systems able to iteratively improve themselves — an approach also called recursive self-improvement. The aim is to automate scientific research and AI research itself, across fields from medicine to materials science.
You can have a self-improving AI where you can point a problem at it and it keeps getting better with time.
— Behnam Neyshabur, co-founder and CEO of Mirendil
The race for compute
The deal fits a broader trend in which young AI labs lock in hardware access years ahead. Rival Recursive Superintelligence earlier signed a compute contract with Amazon worth about $400 million. Mirendil co-founder Harsh Mehta stressed that the startup's models are "really good at working with different workloads and chips", letting it match jobs to available hardware instead of committing to one architecture.
| Lab | Provider | Value |
|---|---|---|
| Mirendil | Google Cloud | over $100M |
| Recursive Superintelligence | Amazon | about $400M |
Why it matters
Compute access has become the main bottleneck for labs working on advanced AI. The deal shows an early-stage startup can contract hardware at a scale once reserved for large players, and that Google Cloud is actively courting such customers. For Google, it signals that the fight with Amazon and Microsoft is not only about chip performance but about entire training environments. For Mirendil, it is a bet: compute scale should turn an ambitious vision of self-improving AI into working systems.
What's next
- The deal is multiyear — Mirendil will scale training on TPUs and Nvidia GPUs as its models grow
- Pressure to deliver rises after the $1B valuation from the seed round closed in late June
- Competition for compute is intensifying — Recursive Superintelligence already secured its Amazon deal
Sources
- TechCrunch — Exclusive: Mirendil inks $100M+ Google Cloud deal to scale self-improving AI
- Google Cloud — Cloud TPU
