Post

Simon Willison
@simonwillison.net

Out of curiosity I decided to try and run the numbers on how much Netflix you can watch for the energy cost of a ChatGPT prompt As far as I can tell it's between 5.1 and 10.2 seconds, depending on which end of the 2019 IEA Netflix energy usage estimate you use simonwillison.net/2025/Nov/29/...

43 reposts217 likes

chris
@chrisecramer.bsky.social

I'll note that .34 wh works out to be a 250 w GPU running for 5 seconds (almost exactly). This sounds low for ChatGPT and also ignores the external costs of transmission, the receiving computer, etc that are rolled up into the Netflix analysis

Simon Willison
@simonwillison.net

I mainly care if the estimate is of the right order of magnitude 12s of Netflix time compared to 10s wouldn't make me feel bad, but if a ChatGPT prompt is actually 2 minutes of Netflix then I'd feel like I'm misleading people in a meaningful way

V
@mirainthedark.bsky.social

LLMs like ChatGPT do get loaded onto more than one GPU - they can't fit on just one. But although that is the case, inference isn't run on just run prompt - it's run in batches. One user isn't hogging up a cluster for their one prompt. They're sharing it with multiple other people.

Simon Willison
@simonwillison.net

The Netflix number is five years old and the ChatGPT number pre-dates GPT-5 by a couple of months so there are plenty of caveats here, but I expect the order of magnitude remains about correct

Simon Willison
@simonwillison.net

Since a lot of people are suspicious of a figure shared by Sam Altman, this independent estimate published in Februwry also came to 0.3 Wh per query epoch.ai/gradient-upd... And Google published "we estimate the median Gemini Apps text prompt uses 0.24 Wh" in August cloud.google.com/blog/product...

cloud.google.comMeasuring the environmental impact of AI inference | Google Cloud BlogA methodology for measuring the energy, emissions, and water impact of Gemini prompts shines a light on the environmental impact of AI inference.
Simon Willison
@simonwillison.net

The big AI labs continue to be infuriatingly opaque about the actual figures for their total electricity and water consumption The best report I have seen from them so far is this one from Mistral in July, but it still left unanswered questions for me simonwillison.net/2025/Jul/22/...

simonwillison.netOur contribution to a global environmental standard for AIMistral have released environmental impact numbers for their largest model, Mistral Large 2, in more detail than I have seen from any of the other large AI labs. The methodology …
David Eccles 🌻🇵🇸7x🩹🛡️
@gringene.org

"I'm particularly interested in understanding the breakdown between training and inference energy costs, since that's a question that comes up in every conversation I see about model energy usage." Me too. Training is typically substantially more computationally intensive than model use.

Greg Langmead
@greg.langmead.info

Appreciate your ongoing effort. It’s a collective resource they are drawing on whether they think of it that way or not. We should refine our methods of holding them accountable.

Chris Krycho
@chriskrycho.com

Thanks for chasing this down—it was my question as well. Altman is just notoriously untrustworthy, and it’s hard to guess whether any given claim is totally unmoored from reality or happens to be one of the times he’s telling the truth. 😭

Simon Willison
@simonwillison.net

That's correct, those numbers do not include the model training cost It's hard to use that as part of a per-query estimate because that then varies depending on how many people use the model during its deployment lifetime - so the training cost per query drops constantly the more time passes

Simon Fondrie-Teitler
@simon.overgrown.garden

I think that there's a pretty decent chance the average electricity usage for Netflix is a decent amount lower than for the average internet traffic too. My guess is these days the electricity usage for Netflix is mostly from the TV.

sawyer
@sawyer-p.me

Yeah Im trying to figure out how inference on a 1T+ param model is the same energy as watching a video, and not even the whole video. You can do video streaming from a raspberry pi to your laptop. That raspberry pi would need a large harddrive and LOT of time to even generate the first token.

Anna McDonald
@jbfletch.bsky.social

This is a really interesting topic, I was curious about print books and it seems the typical paper back uses 4.5 KwH (I searched in vain for the definition of typical but alas could not find it) I declare print books to have a relatively high ROEI!

J. C. Cantwell 🌻
@segfaultvicta.aludel.xyz

video encoding/decoding is wildly more intensive than anyone thinks or wants to believe it is; this is *also* a significant part of why *generative video* is energy-intensive compared to audio/text/images. streaming video and tiktok take up enough energy that text/image LLMs are a rounding error

James Felix Black
@tft.io

This is correct. We did this at iTunes back in the day; had several parallel streams at different data rates/capabilities, and we invested a LOT of energy in the encoding process, because each stream would be decoded thousands of times more than it was encoded.

Phillip Carter
@phillipcarter.dev

Maybe! It could also have gone up with the rise of more and more ultra-HD streaming. Though I don't know how you would decide to count it. From a carbon emissions standpoint, I would say that TV power use should count. But then to apply that to LLMs it'd also have to include device power use.

V
@mirainthedark.bsky.social

This is a really cool consideration and calculation to put things in perspective! Another useful way to frame it using the same numbers is number of prompts per Netflix hour: ~700 for the 5.1 seconds scenario and ~350 for the 10.2 seconds one.

sawyer
@sawyer-p.me

The real question is how I can leverage this info turn off Cursor and watch 4K Netflix all day at the office. Telling my boss they're computationally equivalent.