Frontier AI startup Mirendil has entered a multi-year partnership with Google Cloud to expand the computing infrastructure behind its research into self-accelerating artificial intelligence. The agreement will give Mirendil access to Google Cloud’s AI Hypercomputer, including both Google-designed Tensor Processing Units and NVIDIA’s accelerated computing infrastructure, as the company scales training, inference, and research workloads. TechCrunch reported that the cloud commitment is valued at more than $100 million, highlighting the growing cost of securing large-scale compute capacity for advanced AI development.
Expanding Frontier-Scale Computing Capacity
Mirendil plans to use the new infrastructure across the full model-development cycle, from pre-training and post-training to reinforcement learning and large numbers of parallel research experiments. Google Cloud said the companies worked together on the design and deployment of the environment across compute, storage, networking, and control systems, with managed training clusters intended to simplify the operation of mixed TPU and GPU workloads. Mirendil is already running a cluster based on Google’s TPU v5P chips, while NVIDIA accelerated computing systems are expected to come online as the deployment expands.
Building Self-Accelerating AI Systems
The startup is developing AI systems designed to automate more of the research process itself, including the design, evaluation, and iteration of experiments that are currently heavily dependent on human researchers. Mirendil describes its long-term goal as creating systems capable of taking on increasingly broad parts of the work performed by frontier AI laboratories, allowing research loops to operate faster and with greater autonomy. The company believes that this approach could eventually make advanced AI research tools more accessible to teams working in fields such as medicine, biology, materials science, chemistry, and robotics.
Matching Workloads With Specialized Hardware
A central part of the partnership is Mirendil’s ability to distribute different workloads across multiple types of accelerators rather than relying on a single hardware architecture. The company says its systems are built to optimize the software layer above the chips and assign workloads to the computing resources best suited to them, which could improve efficiency as model training becomes more complex. Google Cloud’s combination of TPUs, NVIDIA infrastructure, high-speed networking, storage, and managed training environments gives Mirendil a flexible platform for running increasingly demanding research workloads at scale.
Strategic Value for Google Cloud and Mirendil
The deal also reflects intensifying competition among major cloud providers to become the infrastructure partners of fast-growing AI laboratories, which require substantial and reliable access to specialized chips. For Mirendil, the agreement expands its compute foundation and provides additional capacity needed to pursue research that depends on running large numbers of experiments continuously and at frontier scale. For Google Cloud, the partnership adds another emerging AI laboratory to its customer base while showcasing its strategy of offering multiple accelerator options within an integrated computing platform.
Mirendil’s partnership with Google Cloud gives the young AI lab a significant infrastructure foundation as it works toward systems that can continuously improve AI research processes. The combination of TPU and NVIDIA computing resources, managed training clusters, and jointly designed infrastructure is intended to reduce operational constraints while supporting larger and more autonomous research loops. If Mirendil can translate that computing capacity into effective self-accelerating AI, the company aims to broaden access to frontier-level research capabilities and accelerate scientific and technological discovery.
Source: TechCrunch