The same AI request emits 8.5 times more CO2 in Texas than in Sweden

jlundberg

Based on a newly released model for calculating this that is open sourced.

armcat

Great writeup! I've done some work on Codecarbon [1] in the past, it's a pretty standard tool for measuring ML/AI workload emissions, it would be awesome to have your findings contributed to that project as well. [1] https://github.com/mlco2/codecarbon

reedf1

We can pretend that international trade and energy price arbitrage doesn't exist to make us feel nice and happy; but the reality is a Watt spent in Sweden increases the marginal cost of energy by the same amount as a Watt spent in Texas. These things are tied together by foreign exchange, imported energy intensive goods, petroleum products and directly imported power. It's time we stop being naive about total energy use.

IshKebab

Yeah I don't understand why these new data centres are all in populated areas. AI isn't latency sensitive. Put them in some remote location surely?

km3r

This is why people in left wing areas need to oppose datacenters bans. Regulate them. Require green energy investments that more than offset their load. Use some of that flowing AI money for the betterment of society.

bob1029

The reasoning effort knob is becoming a really important variable regarding how much impact AI models will have on energy consumption. Medium and max can differ by as much as two orders of magnitude in their token consumption. We also have to consider the amplification of compute demand that occurs by way of the tool calls themselves. Tokens are mostly a constant cost thing, but how they are interpreted can mean the difference between a 5 minute cpu bound operation and a 50uS cached database lookup. At some point, the CPU is going to become the bottleneck again.

alpineman

And eating a single burger results in the equivalent Co2 emissions of 17,000 requests in Sweden and 2,000 requests in Texas. Reducing/stopping meat consumption and replacing it with healthy plant-based tofu, seitan, tempeh etc has a far bigger impact (and as a bonus, living beings don't have to suffer)

wartywhoa23

Worry not, AI people! Greenhouse gas exemptions will be made for the future of work, maths and art, the salvation of the humankind, the genie that is out of the bottle, etc, if only at the expense of taxing your and your dog's fart.

blfr

The comments section reminds me of The Dead Great-Grandfather Test [Science fiction writer Bruce] Sterling was using the term specifically to describe the hairshirt green lifestyle that is driven by eco-anxieties. For hairshirt-green types, life is all about saving water, recycling, composting, reducing eco-footprints and various other behaviors marked by a kind of fearful, non-generative retreat from living. Permanent existential hibernation. Sterling's rule of thumb for spotting acting-dead behaviors is a great one: if it's something your dead great-grandfather can do better than you, it's a case of acting dead. Your dead great-grandfather uses no water or plastic, and is actually recycling himself as we speak, not just his possessions. Try and top that. But acting dead goes beyond hairshirt-green behaviors. While spartan frugality is a virtue, when it becomes the entire purpose of your life, there's a problem. For a portion of the dying American middle class, frugality has turned into a life purpose. https://ribbonfarm.com/2011/12/08/acting-dead-trading-up-and...

skeledrew

Is this site getting a death hug? I got a 502, then 503, then incomplete page load.

arjie

Site is HTTP 503. It does make sense that an nginx 503 would cost much less to serve.

mkgiga

too many things can happen in the ai provider's backend which won't be observable from the outside. for starters how will you even know what hardware and inference backend is being used by RandomAiProvider3284.com? you can't. i looked at your berget-ai/co2-calculator repo and you're expected to provide this yourself. it also uses "measuredResponseTimeSeconds" as a metric even though it's impossible to know how much of that time was spent rate limiting the user during hours of high activity

mepiethree

This is true of ~everything that uses electricity, not just AI https://app.electricitymaps.com

looofooo0

same for Germany which has roughly the same Co2 per kwh.

blaze33

For 2025 the grid carbon intensity in Texas was 363gCO2eq/kWh, while it was 22gCO2eq/kWh in Sweden, so 16,5x more in Texas. Right now it's 427 in Texas vs. 50 in Finland ie. 8,54x more, like in the article (yes I looked the wrong country at first). The reason is half of Texas electricity production is still fossil fuels (gas & coal) while mostly hydro, nuclear and wind in Sweden. So it's not really about ai, the models or the hardware. Fossil fuels make CO2, that's it.

maelito

That's why Mistral is even more important than just being the only competitive european LLM builder : it's on the French grid.

cm2187

Fair comparison. Texas just needs to flood the valleys of its famous mountain chain and collect the melted snow with dams to match Sweden's 40% hydro electricity production.