Mirendil taps AI Hypercomputer TPUs and GPUs for pre- and post-training applications

Google Cloud is announcing that Mirendil is running pre- and post-training workloads on Google Cloud AI Hypercomputer infrastructure that combines TPUs and GPUs. The company is positioned among the high-growth AI startups Google Cloud is highlighting as active users of that stack for model work beyond a single training pass.

On the product side, the setup is not GPU-only or TPU-only. Mirendil is using AI Hypercomputer access that spans both TPUs and GPUs and applying that capacity to pre-training and post-training stages. That split matters because pre-training and post-training put different pressure on compute, memory, and orchestration, so a mixed accelerator path is a concrete deployment choice rather than a single-chip bet.

For engineers and builders, the signal is operational: a startup is treating Google Cloud as the home for both early training and later refinement, not only one-off jobs. Teams building similar pipelines can map their own pre-train and post-train stages onto the same class of shared infrastructure that Google Cloud says major AI labs already use for model training, agent inference, and frontier research.

Market context is the wider Google Cloud AI customer base. Nearly every major AI lab is already on Google Cloud for training, agent inference, and research, and Google Cloud is also pitching itself as the default platform for new high-growth AI startups. Mirendil is being named as part of that second group, which tightens the link between lab-scale usage patterns and startup-scale pre- and post-training work on the same cloud.

What to watch next is how Mirendil’s pre- and post-training mix lands in practice on AI Hypercomputer TPUs and GPUs, and whether more startups in the same cohort publish the same dual-stage pattern. The useful follow-up is concrete: which stages stay on TPUs, which move to GPUs, and how that split shows up in training and refinement workflows rather than in branding alone.