Mirendil signs $100 million-plus Google Cloud deal to build AI that improves itself

Key points

• Mirendil, an AI startup, has signed a deal worth more than $100 million with Google Cloud to access computing power for its research.
• The deal covers access to Google's TPUs (Google's own custom chips) and Nvidia GPUs (specialised processors used to train AI models), giving Mirendil flexibility to run different workloads on different hardware.
• Mirendil raised seed funding at a $1 billion valuation in late June 2025, making the Google deal worth roughly half that total raise.
• The startup is building what researchers call recursive self-improvement, meaning AI that can keep getting better at a task without humans rewriting it each time.
• Mirendil's co-founders previously worked at Anthropic, the AI safety company behind the Claude assistant.

A little-known AI startup called Mirendil has quietly locked in one of the larger cloud computing commitments seen among early-stage AI companies. First reported by TechCrunch, the multi-year deal with Google Cloud is worth more than $100 million.

That figure sits alongside a striking data point: Mirendil raised seed funding at a $1 billion valuation in late June 2025, meaning this single infrastructure contract accounts for roughly half of everything the company raised.

What is Mirendil actually building?

Mirendil is working on self-improving AI, sometimes called recursive self-improvement, which refers to AI systems that can refine their own performance over time without humans updating the code after every step.

Think of it this way. A standard AI tool is fixed at the moment it ships. You can ask it questions, but it does not get better at your specific problem the longer it runs. Mirendil wants to build systems that behave more like a scientist who keeps reading papers, running experiments, and learning. Point the system at Alzheimer's disease research, for example, and it should keep making measurable progress on its own.

"How can we have an AI system that keeps doing research, keeps improving its own knowledge and performance when it comes to Alzheimer's disease?" Mirendil CEO Benham Neyshabur said. "This technology allows us to set goals that are ambitious for AI, and the AI would keep making progress."

Neyshabur and co-founder Harsh Mehta both previously worked at Anthropic, where recursive self-improvement is an active area of study.

Why does this need so much computing power?

Building AI that teaches itself is enormously expensive to run. Training an AI model, the process of feeding it data until it learns patterns, demands huge amounts of specialist computing hardware.

The Google Cloud deal gives Mirendil access to two types of chips: TPUs, Google's own custom processors built specifically for AI training, and Nvidia GPUs, the industry-standard chips that do the heavy number-crunching most AI research relies on. Mehta said the ability to mix and match workloads across chip types should reduce costs for Mirendil and, eventually, for its customers.

Google gets something in return too. Mirendil's software layer is designed to help customers squeeze more performance out of Google's hardware. That makes Mirendil a useful partner for Google to bring to enterprise clients who want cutting-edge AI without building the infrastructure themselves.

What does this mean for ordinary people?

For most readers, nothing changes today. Mirendil has no consumer product. The near-term customers are likely large research institutions and enterprises.

The longer arc matters, though. If self-improving AI works as its proponents claim, it could accelerate research in medicine, biology, and materials science, fields where progress is currently slow and expensive. Faster drug discovery or better understanding of diseases like Alzheimer's would affect everyone eventually.

For now, watch for other early-stage AI startups announcing similarly large cloud commitments. Securing computing capacity has become as important to AI companies as raising cash.