Thank you! Regarding performance, TPU v5e which was benchmarked in the MLPerf results and showed impressive performance/dollar, see https://cloud.google.com/blog/products/compute/performance-p... which has more details. Combined with k8s & GKE, TPU v5e workloads can leverage auto-scaling, for example setting up autoscaling based on traffic, so workloads can scale down rapidly when not in use, increasing cost efficiency.