Mirendil taps Google Cloud AI hypercomputer for research

Mirendil has selected Google Cloud's AI Hypercomputer for its AI development work, with access to both Google TPU systems and Nvidia infrastructure on Google Cloud.

Mirendil, a frontier AI lab, will use the infrastructure for model pre-training and post-training. Google said the setup is intended to support the lab's end-to-end training workflows, including reinforcement learning at scale.

Google said Mirendil chose a combination of TPU and Nvidia systems so it could match workloads to different architectures over time. The arrangement also enabled the company to secure computing resources quickly.

According to details released alongside the announcement, the two companies worked on the design and deployment of the infrastructure across compute, storage, networking and control systems. They also collaborated on a system using managed training clusters in Gemini Enterprise Agent Platform to manage both TPU and GPU environments.

Mirendil is already operating a cluster of TPU v5P chips, with Nvidia-based systems due to be added. The deployment gives the startup access to the two main AI hardware approaches currently used by model developers.

Research focus

Mirendil said its work centres on building AI systems designed to speed up AI research and development itself. That includes handling training processes from initial pre-training through later-stage post-training.

The company is also working on systems that can support reinforcement learning at large scale. Such work has become increasingly dependent on access to large volumes of specialised computing infrastructure, an area where major cloud providers are competing to supply AI labs and startups.

Google said many large AI labs already use its infrastructure for training, inference and research. Adding Mirendil gives it another startup customer in a market where cloud providers are seeking to attract AI companies with access to both proprietary and third-party chips.

That competition has intensified as demand for AI computing has outstripped supply in parts of the market. Access to Nvidia hardware in particular has become a strategic issue for many AI developers, while Google has also promoted its own TPU line as an alternative for certain workloads.

The Mirendil deployment reflects a broader pattern among AI companies seeking to avoid relying on a single chip architecture. By using both TPU and GPU systems, developers can spread workloads according to performance needs, software compatibility and resource availability.

In a statement, Mirendil outlined the reasoning behind its work.

"Progress in AI has been bounded by how fast humans can run the research loop - designing experiments, evaluating results, and iterating. We're building AI systems that can accelerate and improve that loop itself. Expanding on Google Cloud gives us the scale and flexibility to push those systems further and put frontier AI research capabilities in the hands of many more scientists and engineers to run that loop faster and at a greater scale," said Behnam Neyshabur, Co-founder and Chief Executive Officer of Mirendil.