Hi, Tom!

The very core metric of this blogpost is questionable:

> Throughout this post I’ll assume the average ChatGPT query uses 0.3 Wh of energy, about the same as a Google search used in 2009. Here’s a summary of why 0.3 Wh is the most reasonable guess right now.

He's listing a bunch of sources in the summary:

> Sam Altman, CEO of OpenAI, published a blog post in June covering a range of AI updates. In it, he casually mentioned that a standard text query uses 0.34 Wh of electricity. Remember, that’s close to the updated estimate that EpochAI came up with (0.3 Wh),

> Mistral AI, another AI company, conducted an environmental analysis of its LLMs. ... Overall, the impacts were low: just 1 gram of CO₂ per page of text generated

> This article in MIT Technology Review — We did the math on AI’s energy footprint. Here’s the story you haven’t heard — was published in May, and was a good overview of many of the complexities. Some stand-out numbers: it quoted an estimate of around 0.93 Wh (let’s call it 1 Wh) for an average Llama text query response.

And concludes with this:

> The 3 Wh estimate in my previous article was probably too high, and something a little lower — possibly as low as 0.3 Wh — seems possible.

That looks like a bold hand-waving to my taste. The author dropped MIT estimation at all and fully agreed with Sam Altman (very questionable source of information). Mistral AI had relatively small models at a time (around 70B) and can't be used to estimate GPT.

EpochAI uses a lot of indirect measures for their own research and base their estimation on a long ago outdated model GPT-4o with "200 billion total parameters (likely between 100 and 400 billion)." ©. The models are way way larger now, even openweight models hit 1 to 3 trillion parameters and have 1 million token context windows.

Both EpochAI and Mistral AI base their estimations on the per-page basis

> I assume that a typical number of output tokens per query is 500 tokens (~400 words, or roughly a full page of typed text). This is somewhat pessimistic—for example, Chiang et al. found an average response length of 261 tokens in a dataset of chatbot conversations

Agentic AIs do not hold this assumption. LLMs now read huge chunks of files, spawn research subagents, write files with thousands lines of code and each model supports <think></think> blocks where they echo and rephrase information multiple times. One agentic session can easily pass 100k tokens of context and context lookups have quadratic complexity. Plus closed LLMs from top AI labs hide their thinking proccess from the end user.

I can't say if "0.3 Wh of electicity" is a wrong estimation for LLM query, but I don't believe this blogpost estimation because sources of this metric do not look good.

Cheers,
Andrei.

On Sun, Jul 26, 2026 at 11:14:44PM +0200, amindfv--- via ghc-devs wrote:

people who are concerned about, e.g. global warming or loss of drinking water due to LLMs

I would invite people who are concerned about global warming or water
usage of LLMs to explain what they disagree with in Andy Masley's
analysis.  He ultimately concludes that the impact on water and CO2 by
my use of LLMs (all day every working day) is negligible compared to
the impact by other aspects of my daily life:

https://blog.andymasley.com/p/a-cheat-sheet-for-conversations-about?open=false#%C2%A7this-post-in-a-nutshell

Tom
_______________________________________________
ghc-devs mailing list -- ghc-devs@haskell.org
To unsubscribe send an email to ghc-devs-leave@haskell.org