WP has been doing great reporting on this (gift link) wapo.st/45V5OTD

We do a lot of that math in this story from May! www.technologyreview.com/2025/05/20/1... Also - will have more in my newsletter Wednesday :) forms.technologyreview.com/newsletters/...

For all the details: www.technologyreview.com/2025/08/21/1...

Researchers still want more disclosure and more standardized data to compare models, but it's interesting to peek behind the curtains here and get some new details. Who wants to publish their numbers next?

There's a water estimate, too: 0.26 mL, or roughly five drops. That assumes 20% of what's withdrawn is recycled and 80% is consumed. (If you convert Altman's ChatGPT number, that's 0.32 mL, so again, pretty close.)

If you assume Google's emissions are based on the average of the grids where it operates, the total emissions would instead be about three times higher, 0.09 grams per query. (Rough math but you get the idea.)

One more interesting tidbit for my climate folks - the emissions estimate, 0.03 grams of CO2 per query, is only that low because the company is taking credit for all its clean energy purchases.

We also don't get any information here about the total number of queries per day or per month. (Give me Q, I want to know Q!)

The company talks in the report specifically about using the median instead of the average because it sees some outliers that can use much more energy than what it considers a typical query.

Entering a lot of text, getting a long response back, and using reasoning models would all likely push that energy figure higher. So would image and video generation, probably, though this report is specifically about text.

But there are still limitations to what's being revealed here. One note is that this is not an average, but a median—so a query that's picked out of the middle of a range. We don't know what that range is, at all, or how it's distributed.

What's notable about the Google news is that we get a lot more information here. Rather than just dropping a number, the company provided details about what's included—not just energy from the AI chips, but also the host machine's CPU + DRAM energy, cooling systems, power conversion, and so on.

The Google number is also in the ballpark of a figure that Sam Altman shared on his blog in June - he said that a query to ChatGPT uses about 0.34 watt-hours of electricity and 0.000085 gallons of water (inconvenient unit, but whatever) blog.samaltman.com/the-gentle-s...

In our analysis published in May (with a lot of help from the research team behind ML.energy) @jamesmodonnell.bsky.social and I found that prompts to text models ranged from 0.03 watt-hours to 1.9 watt-hours - with a medium model (Meta's 3.1 70B model) hitting pretty close to this Google figure.

The company points out that this is lower than some numbers that often float around (a popular one is 3 watt-hours, that one is quite old at this point and also was an estimate, not a direct measurement).

This rocks - congratulations!! 🥳

OpenAI has said that ChatGPT receives 1 billion messages every day (as of December)

And the original blog post is here blog.samaltman.com/the-gentle-s...

Notably, OpenAI declined to share any info with us when we were putting our story together. We would welcome more transparency, including figures for image and video generation.