[tqi]

> The report estimates that training the latest frontier large language models, such as xAI’s Grok 4, can generate over 72,000 tons of carbon-equivalent emissions.

That seems pretty trivial, relative to 38bn per year globally?

[idoubtit]

The training of one LLM requires as much emissions as 17,000 people over a year. Which, according to the article, is 8 times more than last year, and may be underestimated by a factor 2.

That does not cover the whole usage: the hardware, the bots that collect learning data, the prompts, etc. And there are now many models of this size, and thousands and thousands at smaller sizes. And some of this parameters are increasing.

AI is estimated to emit more than 80e6 tons of CO2-equivalent this year. Much more than whole countries like Austria or Israel. Is that trivial?

[_aavaa_]

How much does that come out to per user or per request? And how does that compare to anything else those people do? Like drive 10 minutes. Or eat a burger.

These numbers keep being put up as large in absolute terms but that’s deceiving for the average person who doesn’t have a way to compare them to something relevant in their lives.

[azakai]

Another way to put it: if training a model cost 72,000 tons of carbon, and it then gets used by 100 million people (typical of major models), the cost per person is 0.00072 tons.

Per the article, the average human uses over 5 tons per year (Americans: 18). Adding 0.00072 to 5 is not really noticeable.

(There is also the cost of inference, of course.)

[jeffbee]

Yeah it's basically nothing despite the fact that xAI seemed to intentionally crank up the carbon intensity for no reason.

Also, hilarious to select 2 major models from 2025 and they're both Grok, almost certainly the least useful, least used, and least interesting of that year.