Comment Re:Onsite generation (Score 1) 46
Grid conditions are highly variable, and if you're in the AI biz, you aren't gonna want to shut down your LLMs for a heat wave.
There's plenty of "AI"-related processing which could be delayed and nobody would notice. Training of new models, for example. You get [access to] a new model a couple days later and you won't even notice, because you get it when you get it already. Google is also sufficiently distributed that they can simply move this processing to another location, since both the queries and the results are very small and there will be no appreciable delay associated with doing the processing far away.