This article suggests a massive oversupply of GPUs for inference, but I don't see any mention of training. I was under the impression that training makes up a significant fraction of OpenAI's/Anthropic's compute budget. Or is that not the case anymore?