MirendilがAI HypercomputerのTPUとGPUを採用、モデルの事前・事後学習に活用Mirendil taps AI Hypercomputer TPUs and GPUs for pre- and post-training applications

Google Cloud has announced that Mirendil, a frontier AI lab focused on accelerating AI development, will adopt its AI Hypercomputer platform to support model training. The move matters because it extends a pattern in which cutting-edge AI research organizations, ranging from established labs to fast-growing startups, are standardizing on Google Cloud infrastructure for their most demanding compute workloads.

According to Google Cloud, nearly every major AI lab already uses its infrastructure in some capacity, whether for training models, running inference for agents, or conducting new frontier research. The company also positions itself as a platform of choice for new, high-growth AI startups that are driving much of the industry's research and innovation. Mirendil's decision fits into this narrative, signaling that the appeal of Google's stack extends beyond the largest incumbents to emerging players building new systems.

The technical core of the announcement is Mirendil's use of a mixed-accelerator setup. The lab plans to combine Google's own TPU AI accelerators with full-stack NVIDIA AI infrastructure running on Google Cloud. This purpose-built infrastructure will support both model pre-training and post-training applications. Google describes Mirendil as a team building new AI systems intended to help accelerate and democratize AI research and development, though specific details about the lab's models or products were not disclosed in the announcement.

AI Hypercomputer is Google Cloud's integrated approach to large-scale AI computing, bundling together compute hardware, high-bandwidth networking, storage, and a software layer designed to run training and inference efficiently at scale. Rather than treating each component separately, the architecture aims to optimize the full system so that customers can extract more performance from expensive accelerators. Offering both TPUs and NVIDIA GPUs within the same environment is a deliberate strategy, giving customers flexibility to match hardware to specific workloads rather than committing to a single accelerator type.

The distinction between pre-training and post-training is central to understanding how modern models are built. Pre-training is the compute-intensive phase in which a model learns general patterns from very large datasets, typically requiring the largest clusters and the longest runs. Post-training covers the subsequent stages, such as fine-tuning, instruction tuning, and reinforcement learning from human or automated feedback, which shape a base model into something useful and aligned for specific tasks. Supporting both phases on the same platform means a lab can keep more of its pipeline within one infrastructure environment.

Google's TPUs are custom application-specific chips developed in-house and refined across several generations, with recent iterations marketed for both training and inference at scale. NVIDIA GPUs, meanwhile, remain the most widely used accelerators across the AI industry, and Google Cloud offers them alongside its own silicon so customers are not forced to choose one ecosystem exclusively. For a lab like Mirendil, running both appears to allow experimentation across hardware while relying on Google's networking and software to tie the pieces together.

The announcement also reflects the broader competitive dynamics among cloud providers. Amazon Web Services, Microsoft Azure, and Google Cloud are all investing heavily in AI-specific infrastructure, including custom silicon, large GPU deployments, and specialized networking, as demand for training capacity continues to strain supply. Google has previously highlighted commitments from other prominent AI organizations, and each new lab that publicly adopts its platform strengthens its positioning in a market where compute availability and cost are decisive factors for startups. Access to large, reliable accelerator capacity is often a prerequisite for competing at the frontier, and cloud partnerships are one way younger labs secure it without building their own data centers.

For Mirendil, the arrangement is likely intended to provide the scale and flexibility needed to iterate quickly on new AI systems. The company's stated ambition to accelerate and democratize AI research suggests an emphasis on making advanced capabilities more broadly accessible, although the announcement does not specify timelines, model sizes, or the exact scope of the deployment.

As a factual boundary, the key points are straightforward: Mirendil is adopting Google Cloud's AI Hypercomputer, using a combination of Google TPUs and NVIDIA GPU infrastructure on Google Cloud, to power model pre-training and post-training. The wider significance lies in what it represents, namely the continued consolidation of frontier AI workloads onto major cloud platforms and Google Cloud's ongoing effort to attract both established labs and rising startups to its purpose-built AI stack.