You can also rent a cloud TPU-v4 pod (https://cloud.google.com/tpu) which 4096 TPUv-4 chips with fast interconnect, amounting to around 1.1 exaflops of compute. It won't be cheap though (excess of 20M$/year I believe).