The creator poses a question about AI's environmental impact, introduces a four-part series based on her research, and then dedicates this first episode to explaining AI's electricity consumption. She presents data on the total electricity use of data centers, estimates AI's portion, discusses future projections, and breaks down the energy cost of a single query, while also cautioning viewers about the self-reported and unaudited nature of the data.
Post caption
I’ve not wanted to speak on this until I did my research and I FINALLY feel like I’m ready to share.
Over the last several weeks my team and I at @catlabs have been deep in the research about AI’s impact on the environment. What we found is that the data isn’t easy to pin down, and there’s a lot we don’t know/that’s not officially reported.
Emotionally… the facts make us feel mixed. More on that later.
Of course, it’s way too much to fit into one video so I’m gonna be doing a four part series. The first episode today is about electricity then it’ll be water then carbon, and then finally what you can do.
If wanna understand in hard FACTS what your impact looks like, follow along.
#ai #datacenters
Transcript
What the fuck is the actual environmental impact of AI? I have seen everything online from it is literally a teaspoon of water to the heat from the data centers is going to melt the ozone layer. So which one is it? My team and I have finally finished a research deep dive on this, and I'm super excited to present the findings to you in my latest series, What the Fuck is the Actual Environmental Impact of AI? These are the episodes. If you want them all, just hit follow. Let's talk about electricity. Okay. So data centers yes approximately tier terawatt hours of electricity in 2024. That's about 1.5% of global electricity. And that's all data better. Netflix, email, cloud storage, and AI all combined. Okay. So of that 1.5%, what percent of that is used by AI? The straight answer, We don't actually know. Not that, like, nobody in humanity no, but we as a public don't know. AI is not metered separately from the rest of a data center. Independent estimates put AI at anywhere better, like, 15% to up to a half of all data center electricity usage, which is a lot. Still, that would only be, like, point two to point 8% of overall global electricity usage. Better that is just where things stand today. What does the trajectory look like? LVNL predicts US data centers reach 325 to 528 terawatt hours by 2028, which would account for 6.7 to 12% of US electricity. Okay. Now let's talk about what your query costs. The current best estimate for a normal text prompt is about point two to point three watt hours. That's the equivalent of running a TV for a couple seconds. Now we have to be careful about where this information comes from. There's some independent research, but a lot of these numbers are self reported by the companies themselves. Google posted a disclosure in August 2025, which is considered to be one of the most detailed so far. That said, it is not an audited or standardized better. And that number doesn't include the amount of energy that it takes to train and create a brand new model. That is a big and separate better only one time cost. And we in the public do not know exactly much electricity it costs to make a new model. We know that training tier three took about 1,287 megawatt hours and gpt four is unofficially estimated around 50,000 megawatt hours. But open a I never confirmed it, so we really don't know. If you're still here, that's amazing. But we're gonna continue the conversation in part two, which is gonna be focused all about water. Follow for more, and I'll see you there.
