AI company Mirendil has signed a multi-year deal with Google Cloud worth more than $100 million to get the computing power it needs for its self-improving AI research, as TechCrunch reported on August 6, 2026.

The deal, confirmed by Mirendil co-founder and CEO Benham Neyshabur, gives the San Francisco-based startup access to Google's custom AI chips (called TPUs), Nvidia's graphics chips (GPUs), and managed training clusters. Having both chip types matters because it means Mirendil is not relying on a single supplier for the heavy, ongoing computing work that AI training requires. Google Cloud can offer both its own chips and Nvidia's under one contract, which gives Mirendil flexibility.

Mirendil's co-founders, Neyshabur and Harsh Mehta, previously worked at Anthropic, a well-known AI company, before leaving to start their own. They raised seed funding at a $1 billion valuation in late June 2026, roughly double the value of the Google Cloud deal. That valuation, achieved within weeks of the founders leaving Anthropic, signals that investors are eager to back Mirendil's specific approach. The Google Cloud commitment is significant but not make-or-break money; it is the startup buying computing infrastructure using funds it already has.

On Google's side, the deal is handled by Amin Vahdat, a senior vice president and chief technologist for AI and infrastructure. His involvement indicates this is a high-priority arrangement for Google Cloud, not a standard startup program. For Google, securing a well-funded AI lab founded by former Anthropic employees is a competitive move. Cloud providers are competing to be the default computing platform for the next generation of advanced AI development.

Mirendil's technical goal is what gives the deal real weight. The company wants its self-improving AI to eventually do the work of an entire AI research lab. The platform is designed to let scientists create their own AI, and the self-improvement process will use reinforcement learning sandboxes. These are controlled environments where AI models can be tested, compared, and improved before being put to real use.

The idea of self-improving AI, where a system gradually gets better on its own through automated trial and error, is not new as a research concept. What is new is the amount of computing power being dedicated to it. A $100 million-plus cloud deal gives Mirendil the capacity to run the large-scale experiments this approach requires. Reinforcement learning sandboxes consume enormous computing resources by nature, because each round of improvement involves training, testing, and comparing potentially thousands of AI model variations. Without access to chip clusters at this scale, the process slows down so much that self-improvement becomes a theoretical exercise rather than a practical tool.

In the competitive landscape, Mirendil's approach differs from other leading AI companies. Rather than building one flagship AI model to compete with Anthropic, OpenAI, or Google DeepMind, Mirendil is building tools for the process of improving AI itself, and making those tools available to scientists who want to build AI for specific fields. A rough comparison would be building an automated research assembly line rather than a single researcher. If the self-improvement process works at scale, the cost of creating capable AI for new scientific fields could drop significantly.

The risk is that self-improving AI is still one of the hardest problems in the field. Figuring out how to automatically select the best models, design effective feedback systems, and prevent the AI from optimizing for the wrong thing are open research questions, not solved engineering. The computing budget Mirendil now has could lead to answers, or it could reveal just how difficult the problem really is. Either outcome would be informative, but only one leads to a viable business on the timeline that a $1 billion valuation implies.

For Google Cloud, the deal is clearly positive: a well-funded customer with strong AI credentials and a research agenda that requires exactly the kind of flexible, high-end computing Google is built to provide. The deeper question is whether Mirendil's bet on self-improvement as the key to progress will pay off, or whether, like other ambitious AI automation efforts before it, the difficulty of the problem outpaces even a nine-figure computing budget.