Anthropic seems more comfortable using TPUs for overflow capacity. The recent Claude degradation was largely due to a bug from implementation differences with T

Anthropic seems more comfortable using TPUs for overflow capacity. The recent Claude degradation was largely due to a bug from implementation differences with TPUs and from their writeup we got some idea of their mix between Nvidia and TPU for inference. I'm not sure if OpenAI has been willing to deploy weights to Google infrastructure.