bringing this out of the comments to chat a little more about it (thanks, @thicccatbutts for letting me use this as an opportunity to practice some of this whole 'avoiding conspiracy thinking' stuff) and here's the article you linked.
It's interesting immediately to me because the article itself seems to be quite supportive of the idea that AI doesn't use that much energy, but to be fair your comment was specifically about my claim that Youtube uses more energy than AI does.
It is good to have the numbers, though, since that's how I can best reevaluate my own thoughts and expectations. I do wish I had even more specific sources outside of the article, since the main thing they reference to show off their numbers appears to be a datacenter company? Which, to be fair, would be the only way to get those specific kinds of numbers, but it also does feel like the sort of thing that would be the most biased.
Regardless! I'll work with these numbers for now. And they have a bunch of specifics, which is nice!
YouTube or Netflix, 1 hour (HD)
~0.12 kWh → 42 g CO₂
Text-to-video generation, 6–10 seconds
~0.05 kWh → 17.5 g CO₂
AI image generation, 1 image
~0.003 kWh → 1 g CO₂
Google search or AI chatbot prompt
~0.0003 kWh → 0.105 g CO₂
To just make it even more even and understandable, I've done a little bit more of the math.
1 Hour of Youtube = 15-24 seconds of video generation = 42 image prompts = 420 google searches/chatbot prompts
So! On the face of it, that *does* seem like quite a lot. Not even 30 seconds of rendering a video and you've already exceeded the energy cost of watching Youtube for an hour.
But I think I have a few counterpoints?
1: The amount of people rendering AI videos and the amount of people watching Youtube are wildly different.
Like, this 8 hour video of someone 100%ing every Donkey Kong Country Game (to just pull an example video off my timeline) has 48k views, and it was only posted a day ago.
To be fair (and I'm gonna keep doing this where I sort of refute my own point because I think that's kind of part of exploring this and following my train of thought) I'm pretty sure I count as one of those views and I only saw about half an hour of it. So maybe it's better to round that down to about 1 hour of watch time per view. And this is just one of thousands or millions of videos on youtube. But honestly maybe there's a better source for how much youtube people watch, someone has to be keeping track of that, right? Let me look it up.
Well, that's a lot, can I verify that?
Going down a chain of sources and I find this article that appears to be making the base claims, and this seems to be the whole thing they do, but I also can't find the actual research and at the bottom they have this disclaimer
So I really just have no actual idea how much people are watching youtube, but I would assume it's somewhere near that scale, probably.
Now, my expectation is that there's not that many people making AI videos, but I should go check now that I'm sitting at my computer and can actually do that!
Buuut it's also pretty hard to find numbers on that, it turns out. Here's the best one I found from this artlcle?
The problem with both of these articles is that I would not be hard-pressed either way to assume they're inflating their numbers to make it look more impressive. But this is what we have to work with, so it's what I'll use.
The article also says how long the average video was, which, how convenient, is about 10-25 seconds! That's roughly about 1 hour of Youtube. So, with 1 point... I'm gonna round up to 5 just to be safe? 1.5 billion videos generated in a month by Grok, that compares to Youtube's 1 billion hours of video a day by... about 1:20.
Which is honestly way more than I expected! I don't use AI, so I don't always necessarily have a good scope for how many people are using it, but I guess it's a decent amount of people after all. (Grok makes sense, since it's literally on Twitter and easy to access, of course people would be using it.)
So, more than I expected, but I'm still not really convinced that it's all that much. I guess that brings me to the second comparison I'd like to make, which is
2: Rendering *non* AI videos. This is actually something I'm much more familiar with because I've dabbled in 3D animation quite a bit!
It's actually very surprising to me that rendering an AI video appears to be WAY more efficient than rendering AI images individually, because all video is is a bunch of individual images lined up in a row. Assuming 24 frames per second, 10 seconds of video is 240 images, which only takes up as much energy as 42 image prompts. I guess since a lot of it is rendering something just slightly different, that cuts down on the cost a lot!
Now, to compare it to non-AI videos... rendering a single frame of an animation can take anywhere from five seconds to five minutes, depending on how intensive the modeling and effects are. For my computer at home, it generally took about 2 minutes. so I'm gonna run with that!
Rendering animation is also *VERY* power intensive. You can certainly do it at home, but any professional animation studio will generally either rent out space in renderfarms or have their own (that is, have a datacenter with a bunch of computers/servers running in tandem to render out the animation quicker.) This article says that one mid-range laptop can use about 600-800w per hour while rendering.
To create 240 frames of video, (or 10 seconds) I would need to have my computer running hot for about 8 hours. That would be ~5.6kW of energy. Which compared to an AI rendered video (0.05 kW) is about a 1/112 difference. You could render about 18 minutes of AI video for the cost of rendering 10 seconds of non-AI video, which, for the sake of an environmental/energy usage cost, actually seems to say that we should be using AI more often instead of making things by hand.
Not that I'm saying that--I really want to animate things by hand! I enjoy it!
And I think all of it comes back to the crux of the matter (sorry, I know it's been a wall of text, but bear with me) is that humans have a really hard time figuring out what is important to be worried about, and that's why it's often important to spend a little extra time asking why people are angry and seeing if it's reasonable. The thing you said at the beginning in response to my comment ("If golf courses were popping up around the country fast enough to be disruptive to the local infrastructure I think people would also be mad about that") is sorta exactly the problem! People/society at large being mad or not mad about something isn't actually proof of whether there's a problem or not. Thousands of people were and are very vocally and prominently mad about vaccines, but that doesn't actually mean they're bad!
And people are mad about golf courses--as they should be! We just don't hear about it as often. There are 26 more golf courses set to be built in 2026, and golf courses in the US use about 738 billion gallons per year, compared to the water usage from data centers in the US, which is about 174 billion gallons of water per year, and that's *every* data center--tumblr, youtube, Ao3, cooking blogs, render farms, and yes, AI, which makes a up a small portion of that. (Though is also about water usage, rather than energy directly, to be fair.)
In the end, there are a lot of things to think about, and we don't always have time to consider it all, so we just have to do our best. How much does this account for AI training? Youtube uploading? Storage of videos, the lawnmowers used to cut golf course grass, ect. Maybe you'll look at these numbers and have a different reaction than I do! But I hope at least it was an interesting read.
