“Google Cloud's infrastructure strategy involves separating training and inference workloads at the silicon level to avoid performance bottlenecks.”