26MW is a fraction of the primary power consumed by a single passenger aircraft, by the way. It is an absolutely trivial energy input.
Figures I can find suggest that a 737 uses approximately 7MW to stay aloft.
So a couple things I learned -- I think it's still a notable amount of power, enough to power ~6,000 homes for a year just to train a single model. But also, I learned that planes use a whole lot more power than I thought!
Training a single model is essentially consuming one plane-year's worth of power, or 3-4 flights continuously while it trains. I had no idea planes used so much energy.
But also, I bet most of these companies aren't training one model and calling it done. There's probably 1s or 10s of models being trained per year per company. That's a material amount of energy use. If we could power tens or hundreds of thousands of homes, that isn't 'trivial' energy input.
I think it's useful to put it into context next to other things we take for granted, but I don't think it's fair to diminish it as nothing either.
"The Dongfang Electric Corporation's 26 MW offshore wind turbine is the largest in the world, surpassing previous models like the Mingyang 20 MW turbine. This turbine's larger size and capacity enable it to generate about 100 GWh of electricity annually, potentially powering 55,000 Chinese homes or 9,200 American homes."
Edit: more info here, https://www.bloomberg.com/graphics/2025-ai-impacts-data-cent...
Your claim: 50-100 liters per second to cool a 26MW workload.
Actual water consumption, according to Google annual report: 730 liters per second, globally, for an average 3GW load.