Like my original point, "training a model" does not inherently involve a GPT-tier amount of electricity, because most AI models aren't eleventy-billion parameter behemoths that require thousands of GPU-aeons to train.
My original point was that "burning energy" and "using electricity" are not inherently carbon-producing Bad Things (otherwise nobody would care about renewables). Electricity Good. Combustion Bad. See also: vehicles.
The second point I was making was that, regardless of the energy required to train this particular model in question, it almost certainly pales in comparison to the literal tons of jet fuel burned per hour in the 70 test flights the article mentions.
Considering that they don't do one thing at a time (like Google Cloud, and not running at capacity because scaling, etc, etc, etc) - I'd also bet selling ads doesn't use a whole DC, either.
Expanding on my other comment (Ctrl+F "IEA"), this seems like knocking out some common-sense, low-hanging fruit: if 35% of aviation's contribution to climate change is because of contrails and we can reduce the number of contrails simply by flying at a different altitude, then we can solve a big portion of a major category of contributor.
Unless we use up entire oilfields of energy "using electricity" to "train models", of course.