Post

I’d have to check the math on this but I’d guess nearly half the queries are below the median, even if yours happen to be above. So it balances out.

So, depending on where you are, $0.20-$0.50 a day of energy usage. That doesn’t seem too bad for such a professionally useful tool.

It is when it's away from someone else. Energy and water are not inexhaustible resources *especially* at a local scale, and they're not measured in dollars. (As part of a wildly unethical technology but I'm guessing we're ignoring that part in this thread.)

Well, they are, and they aren't, and when you measure how much it affects someone else elasticity is what matters. But fair play, let's try to get a feel of how much this is--order of magnitude, anyway. So using his numbers, that's about 1.25KWh total usage...

So what are you doing about golf courses?

energy and water are both measured in dollars where i'm at. i don't always remember my ccf to gallons to liters or kWh to MJ conversions, but i do know my monthly bills

cc @andymasley.bsky.social @simonwillison.net might be interesting for you guys

(1/2) AI's very fast-moving and coding agents didn't exist when I started. I can only really cover so much at once. I wish people saying "it's deceptive to not also cover X" would consider just how many individual angles you need to cover.

(2/2) I would be a little surprised if the audience talking about AI and the environment were also regularly using coding agents, they've seemed pretty distinct.

They pick the median probably because it makes them look better. In this case it's probably better to use an average...

Who's "picking" anything here? He's accurately saying he's a far outlier, using way more than typical, and showing how much that is.

I mean the discourse using the median. Mean is much more appropriate to discuss impacts like this because it includes rather than excludes far outliers.

hey, nice work! what does your (cached) input:output ratio look like? I don't strongly vouch for our estimates of input/attention costs still being relevant today so would be useful to know how sensitive this is to inputs vs outputs

another sanity check: you spend $15-20/day. if you extrapolate to Anthropic's ~10B revenue this is 85 MW for Anthropic inference. disclosure: Claude did the math here but it looks right. This actually looks too low. but because of subs their actual revenue per tok is less, so actual tokens higher?

all Anthropic data centers are probably >1 GW: Anthropic's single largest is ~700 MW. And OpenAI says they have ~2 GW in total. This is split between inference and training/research but I'm quite sure inference is >8%. There's also consumption vs capacity which shouldn't be a huge difference.

Very much appreciated your blog post when putting this one together. Thank you. That ratio is very, very input-heavy, so would likely be quite sensitive to those estimates.

Thanks for sharing & it's super interesting to see from someone not at the 'median'! And not to discount what you found but it seems this might be more about the folks who use these tools frequently. And maybe that's worth exploring by itself since more people are using it like you (writing code)

The one that got me was when you said your "median session" was ~150x a "typical query" and then your "typical day" used equivalent of ~4500 "typical queries" Does that mean you have ~30 sessions per day? Let's say they take a conservative 5 minutes of wait per session— that's ~2.5hrs of waiting?

Yep, something like 30. The "2.5hrs of waiting" is where the 2-3 agents at a time comes in, multitasking between agents, but I also wonder whether subagents working in parallel within a session might also drive that "actual time spent waiting" down?

Nice post! You might also be interested in my attempt to get some figures on inference energy usage, but coming from the perspective of the concrete data available in the InferenceMAX benchmarks. muxup.com/2026q1/per-q... Many many provisos and limitations of course

Thanks for the analysis! This does not include model training, right? Do we have a notion of how much the training represents in terms of energy consumption?

It does not, no. Mistral AI is the closest-to-frontier org that has released a comprehensive environmental impact report including data along those lines (that I know of) mistral.ai/news/our-con...

If I read correcly, Mistral says that the energy footprint of training is about 18 billion times a standard response (20.4e6/1.14e-3). How many queries does a model serve in its lifetime? w. 1 bn: training footprit/query ~18 times higher than the per-query usage. w. 10 bilion ~1.8 times higher