That’s the problem of reference. Your individual queries might not consume much - especially when compared to the training - but the more people use it, the more the whole consumption is. At some point running those models will consume more than training them
I recently tried to calculate this for my company. I wouldn’t call it negligible, but the impact of all video calls turned out to be much greater than the impact of AI.
Training GPTs takes a lot of water and energy, running them doesn’t take massive amounts.
Yes it does. Not as much as training is a stratospheric bar…
https://doi.org/10.1145%2F3630106.3658542
That does less than nothing to disprove my point…
that’s old data. inference uses more than training now, since usage has gone up significantly. they traded places in march or april 2025.
That’s the problem of reference. Your individual queries might not consume much - especially when compared to the training - but the more people use it, the more the whole consumption is. At some point running those models will consume more than training them
we passed that point last year, yes.
I recently tried to calculate this for my company. I wouldn’t call it negligible, but the impact of all video calls turned out to be much greater than the impact of AI.
Of course that’s going to produce a heavier load on your part of the infrastructure… That stuff is running locally.
Also, it’s pretty easy to have effective calls and turn off the damn video. Most people don’t need to stare at everyone staring at their computers.