Mirendil signs $100 million-plus Google Cloud deal for self-improving AI research

AI research company Mirendil has signed a multiyear Google Cloud partnership to secure computing capacity for work on self-improving AI systems. Chief executive and co-founder Behnam Neyshabur told TechCrunch that the agreement is worth more than $100 million. The deal gives Mirendil access to Google's tensor processing units, Nvidia graphics processors and managed training clusters. It is an infrastructure commitment for research, not the release of a public model, and the distinction is important: the announcement establishes a large compute relationship while the system Mirendil ultimately hopes to build remains under development.

The mixed hardware is central to the partnership. Different stages and types of AI workloads can favor different accelerators, so access to both TPUs and Nvidia GPUs gives the lab options when assigning training and research tasks. Mirendil co-founder Harsh Mehta said the challenge is increasingly to match workloads with the appropriate chips. Managed training clusters add a systems layer around those accelerators, allowing the company to focus not only on raw chip access but also on coordinating large jobs. Mirendil argues that this flexibility can lower costs for its own work and eventually for customers using its systems.

Mirendil describes its research goal as AI that can iteratively improve its knowledge and performance. Neyshabur compared the idea with the way scientists learn a new domain, accumulate expertise and become more effective over time. The company believes such systems could automate parts of scientific and AI research and contribute to work in medicine, biology and materials science. Those are ambitions rather than demonstrated outcomes from the cloud agreement. The deal supplies resources for pursuing them; it does not establish that Mirendil has already built a system capable of independently making discoveries in those fields.

The scale of the commitment illustrates how access to compute has become a strategic input for young AI labs. Mirendil is securing capacity through a multiyear relationship rather than relying only on short-term availability. For Google Cloud, the agreement places its infrastructure beneath a company pursuing frontier research and gives it a partner developing a software and systems layer over Google's hardware. Amin Vahdat, Google's senior vice president and chief technologist for AI and infrastructure, framed current AI progress as a problem of orchestrating whole systems as well as improving individual chips.

That systems emphasis aligns with Mirendil's pitch. Neyshabur said the company's software can help customers obtain more from Google's hardware, while Google can eventually offer the resulting technology to enterprise users if the research matures. The relationship is therefore more than a straightforward hardware rental in the way both companies describe it. Google supplies varied accelerators and managed clusters; Mirendil works on methods for allocating workloads and improving systems over repeated cycles. Still, any future enterprise offering remains prospective, and neither company announced general availability of a Mirendil model on August 6.

Mirendil says training self-improving AI requires enormous amounts of computing power. Mehta said the task increasingly involves matching the right workloads to the right hardware, making access to several kinds of chips valuable to the research. The Google partnership addresses that requirement by providing access to TPUs, Nvidia GPUs and managed training clusters. Mirendil argues that this flexibility will let it mix and match workloads with suitable accelerators and lower costs for both its own work and customers using its systems.

For now, the concrete development is the signed partnership and the resources it makes available. Mirendil has a multiyear path to Google TPUs, Nvidia GPUs and managed training clusters under an agreement its chief executive values above $100 million. What follows remains open: how the lab will measure iterative improvement, which research tasks its systems can handle and when any resulting product might reach customers. The announcement is significant because it equips a well-funded research program to pursue those questions at scale, while stopping short of claiming that a public self-improving AI system has already arrived.