Hmm, after running an LLM locally I'm not too surprised about low energy use while querying (but even then, 3Wh is very low, I was getting like 5 Wh per query while running DeepSeek locally, that said datacenter GPUs will be more efficient for this). But also, this doesn't go into great depths about the impact of training.

By your numbers 1 prompt generates 3g of CO2

ChatGPT says they handle 2.5billion prompts per day.

That's 7,500 tonnes of CO2 per day from ChatGPT alone. The vast majority of those are completely pointless queries.

Your point is spurious at best and misleading at worst.

1 Bitcoin transaction = 645kg of CO2

I think it's interesting that this very much depends on the coin. This article compares Bitcoin to Monero, and here are the figures (per transaction per year):

• Bitcoin - ~95 TWh per year
• Monero - ~1.5 TWh per year

So Monero is ~63x more efficient than Bitcoin at transactions, and that's with all the overhead Monero bakes in to its transactions to maintain privacy (each transaction generates a bunch of fake transactions to mislead snoops, and it's notoriously ASIC/GPU inefficient).

I'm guessing the same is true for different AI models, some models will be a lot more wasteful than others in their training and queries.

I agree, the opposition to LLMs is misplaced. There are legitimate reasons to dislike it, and there are certainly policy changes we should make, but attacking it for energy use in the weakest way to oppose it. If LLMs ran on 100% green energy, would you still oppose them? Probably, so figure out why you oppose it and attack LLMs over that, and perhaps policy can alleviate the worst of those concerns.

Upvoted for controversial post...

This whole analysis is based on the cost of queries, which consume, in fact, a small amount of power.

You (or OOP) also acknowledge that the massive, obscene, megalomaniac power consumption of AI comes from building up the models themselves.

And then you proceed to ignore that the first one is the driver for the second. This is asinine. It's exactly like saying that the CO2 emissions of fossil fuels is the cost of running the pumps at the station.

Here you go, according to the same logic:

the beef number comes from poore-nemecek 2018, which is not good science.

So all the data centers they've covered the country with aren't burning electricity and water and generating waste?

The shipping industry emits a billion tons of CO2e per year. Training a model emits... maybe a thousand? An impact that could be offset by reducing Chinese imports by 0.0001%. Or arbitrarily limited by strong-arming the very few companies involved. DeepSeek knocked off a few orders of magnitude and R1 seems to work, as well as any of these things work.

But some people don't really give a shit about the electricity involved - it's just a negative for them to latch onto.

Now, it is a problem locally, where datacenters turn power straight into heat, as fast as they can manage. Anything like a tiny sliver of the global shipping industry becomes noticeable when it's concentrated in one building within range of a commute.

What I think this fails to account for is the most energy intensive part of the models is not interacting with them but training them, which uses exponentially more energy.

Edit: I read over your source more carefully and I don't think the argument of "AI might be using 10 times less energy than we think" is an actual argument of any sort. I might have a golden pony who radiates pure love and shits chocolate donuts, prove I don't.

Here's facts, data centers are ramping up to train and operate AI and making water near their locations inaccessible source

Not to mention Musk's gas turbines source

The carbon output from any form of large scale industry is going to have an impact but a larger issue for me is depletion of resources and a frenzied mania surrounding some imaginary 'race' to beat others at a technology which is yet to turn a profit for anyone but Nvidia