Then he can spread that energy usage out over 20 years, so not to burdent the infrastructure and drive up prices for everyone. That way he needs a much smaller data canter too.

A human might be using something like 300 times more energy, generating the same number of tokens. But those tokens are higher quality. The LLM will be generating much more text for the prompter, than a human assistant would, for the result to be useful. The LLM will also be issued allot of low value prompts, driving up it's over all cost of operation. I've seen people use it instead of a calculator. It's also generating useless text with every google search.

Worth noting that humans deriving energy from food, are orders of magnitude more efficient than a server deriving it's energy from coal/gas/nuclear power. Poweline transmission losses alone probably make up the difference between human and LLM.

Unless he can make LLMs run on brains in jars, powered by sugar, we are comparing apples and potatoes here.