Not at Google but currently using Jax to leverage TPUs, because AWS's GPU pricing is eye-gougingly expensive. For the lower-end A10 GPUs, the price-per-gpu for a 4 GPU machine is 1.5x the price for a 1 GPU machine, and the price-per-gpu for a 8 GPU machine is 2x the price of a 1 GPU machine! If you want a A100 or H100, the only option is renting an 8 GPU instance. With properly TPU-optimised code you get something like 30-50% cost saving on GCP TPUs compared to AWS (and I say that as someone who otherwise doesn't like Google as a company and would prefer to avoid GCP if there wasn't such a significant cost advantage).