Doing calculations like these is a really bad idea because they muddy the waters, putting the focus on disputable numbers - unless you're willing to do a proper study and really dive into it. It's extremely tiring seeing all these well-meaning people, including possibly this author (unless they're writing this for clout), making this mistake time and time again. You're actively not helping.

There's a much better heuristic. Just look at the contracts that the model providers (MS, Google et al) have recently, suddenly established to buy lots of electricity, spin up nuclear power plants and so on. That's the giveaway and one that doesn't need any calculations. Plus the hundreds of billions being invested in datacentres purely for these models.

You see it in this very comment section, lots of naysayers purely based off the provided, utterly meaningless numbers, when the above facts say it all and renders everything else moot. If the training and inference wouldn't cost insane amounts of energy, they wouldn't be hastily taking up these contracts and investing such obscene amounts in datacentres. That's all you need to know.