Post · bonfire.cafe

Coming at this another way: MIT cites 50 GWh for training GPT-4. Depending on US zip code, that's going to be somewhere between 11-20 metric tons of CO2 equivalence. That's 1000x less than the cited Mistral number.

So we have to conclude that training is bad, but given current usage rates, inference will blow away that number in short order, so don't let anyone cite per-inference costs as a stat in LLMs' favor.

https://www.technologyreview.com/2025/05/20/1116327/ai-energy-usage-climate-footprint-big-tech/

MIT Technology Review

We did the math on AI’s energy footprint. Here’s the story you haven’t heard.

The emissions from individual AI text, image, and video queries seem small—until you add up what the industry isn’t tracking and consider where it’s heading next.

Tim Cowlishaw

@mistertim@assemblag.es · last week

@mttaggart m'colleague @fershad did a bit of a survey of this - there's precious few, fwiw: https://carbontxt.org/ai-model-cards

https://carbontxt.org/ai-model-cards

Kyle Hughes

@kyle@mister.computer · last week

@mttaggart This would really give up the game for the Chinese models because they are being distilled from American frontier models.

Taggart :ifin:

@mttaggart@infosec.exchange · last week

Oh yeah! Mistral did publish some data, for what it's worth:

https://mistral.ai/news/our-contribution-to-a-global-environmental-standard-for-ai

Our contribution to a global environmental standard for AI | Mistral AI

Taggart :ifin:

@mttaggart@infosec.exchange · last week

Okay so if I did this right with the EPA's converter, training/using (marginal) Mistral was mind-blowingly bad. Please someone check my math. The stated CO2e was 20.4 kilotons.

A screenshot of a Greenhouse Gas Equivalencies Calculator showing that 20,400 Metric Tons of Carbon Dioxide equivalent (CO2e) is comparable to various real-world scenarios. This includes greenhouse gas emissions from 4,758 gasoline-powered passenger vehicles or 18,016 electric-powered passenger vehicles driven for one year, and 51,949,815 miles driven by an average gasoline-powered passenger vehicle. It is also equivalent to CO2 emissions from 2,295,488 gallons of gasoline consumed, 2,003,929 gallons of diesel consumed, 22,660,624 pounds of coal burned, 270 tanker trucks' worth of gasoline, 2,740 homes' energy use for one year, 4,251 homes' electricity use for one year, 113 railcars' worth of coal burned, 47,230 barrels of oil consumed, 937,157 propane cylinders used for home barbeques, 0.005 coal-fired power plants in one year, 0.053 natural gas-fired power plants in one year, or 1,649,274,178 smartphones charged.

Taggart :ifin:

@mttaggart@infosec.exchange · last week

Seems like they have combined training and 18 months of inference. So the question is how much of this number was pure inference, and what was training.

Taggart :ifin:

@mttaggart@infosec.exchange · last week

So we have to conclude that training is bad, but given current usage rates, inference will blow away that number in short order, so don't let anyone cite per-inference costs as a stat in LLMs' favor.

https://www.technologyreview.com/2025/05/20/1116327/ai-energy-usage-climate-footprint-big-tech/

MIT Technology Review

We did the math on AI’s energy footprint. Here’s the story you haven’t heard.

The emissions from individual AI text, image, and video queries seem small—until you add up what the industry isn’t tracking and consider where it’s heading next.

1 more replies

PhDog

@dogfox@kpop.social · last week

It's very weird to bundle them like that. Looks like they want t hide the worst.

The only way to sus it out I can see is to take their 1.14g of Co2 per 400 tokens and try to estimate how many tokens what fraction of the 20 kilotons buys you. I guess the hing to do would be to dig up user/revenue numbers to try to estimate that.

Revenue ~30-60 million USD/yr:
https://electroiq.com/stats/mistral-ai-statistics/

Users pay ~$20/mo. so, ~4-5 million users? How many tokens you think they ran through in 18 mo.s?

@mttaggart

Electro IQ

Mistral AI Statistics By Revenue And Facts (2025)

Mistral AI Statistics - Mistral's valuation jumped from US$260 million in June 2023 to US$6.2 billion in June 2024.

David Huggins-Daines

@dhd6@jasette.facil.services · last week

@mttaggart interesting... Deleted my previous reply since I assume their calculation of CO2e takes the electricity mix into account, but it's still surprising given that the French grid is 67% nuclear and 25% renewable. Did they train the models somewhere else?

David Huggins-Daines

@dhd6@jasette.facil.services · last week

@mttaggart not really because they used French electricity which is 67% nuclear and about 25% renewable

David Huggins-Daines

@dhd6@jasette.facil.services · last week

@mttaggart Ai2 did this for OLMO2, see the environmental impact section: https://arxiv.org/pdf/2501.00656

https://arxiv.org/pdf/2501.00656