Which Google Cloud accelerator is designed for this workload?

TPUs are Google's custom-designed chips for large matrix operations, and Google's own suitability list names this profile: models dominated by matrix computations, runs measured in weeks or months, and large effective batch sizes. GPUs are the right answer when a model carries many custom operations that have to run outside the accelerator's optimized path - a real distinction, and not the one this workload turns on. CPUs will run the job and would take orders of magnitude longer, because nothing about general-purpose compute is built for matrix maths at this scale. Provisioned Throughput is the one to recognise and rule out: it reserves capacity to guarantee latency on inference, so it buys nothing at all for a training run.