You can cherry pick the numbers in either direction technically - you can operate an LLM without a GPU to make the AI power usage really low, and you can play a game with only basic 2D visuals that has low power usage.

However, the ideal in both cases will fully utilize a high powered GPU, so it’s a wash. During the day my GPU is answering prompts using the Qwen 3.6 LLM, and at night it’s ray-tracing in Cyberpunk 2077. In both cases it’s using a ton of power, but in only one of those two scenarios is it actually producing anything of value.