[utopiah]

Does your comparison include training, data center building, GPUs productions, etc or solely inference? (genuine question I don't know the total cost for e.g. Sora2, only inference which AFAIK is significant yet pale in comparison to everything upstream)

[pbmonster]

No, that's one reason why there's at least an order of magnitude wiggle room there. I just took the first number for J/Token I found on arxiv from 2025. Choosing the exact model and hardware it runs on is also making a large difference (probably larger than your one-time upfront costs, since those are needed only once and spread out across years of inference).

My point is mobility, especially commercial flight, is extremely energy intense and the average westerner will burn much more resources here than on AI use. People get mad at the energy and water use of AI, and they aren't wrong, but right now it really is only a drop in the ocean of energy and water we're wasting anyways.

[utopiah]

> right now it really is only a drop in the ocean of energy and water we're wasting anyways.

That's not what I heard. Maybe it was in 2024 but now data centers have their own categories in energy consumption whereas until now it was "others". I think we need to update our collective understanding in terms of actual energy consumed. It was all fun & games until recently and slop was kind of harmless consequence ecologically speaking but from what I can tell in terms of energy, water, etc it is not negligible anymore.

[pbmonster]

Probably just a matter of perspective. It's a few hundreds of TWh per year in 2025 - that's huge, and it's growing quickly. But again, that's still only a small fraction of a percent of total human primary energy consumption during the same time.

[mlrtime]

You could say the same about the airplane, does the CO2 emissions that the airline states for my seat include building the plane, the R/D, training the pilot.

[utopiah]

Sure and I do, it's LCA https://en.wikipedia.org/wiki/Life-cycle_assessment the problem IMHO being that the AI hype entire ecosystem is literally hiding everything it can about this behind the veil of giving information to competitors. We have CO2eq on model cards but we don't have much datapoints on proprietary models running on Azure cloud or wherever. At best we infer from some research papers that are close enough but we don't know for the most popular models and that's quite problematic. The car industry did everything it could too, e.g. Volkswagen scandal so let's not repeat that.