“LLM companies are better at running inference on GPUs than traditional hyperscalers are.”