[shawndrost]

You have this extremely wrong. TLDR: LLMs singlehandedly reversed the secular decline in US power emissions, which were the only reason for climate optimism.

The story of US power sector emissions was a good story. Emissions appeared to be in secular decline from 2005 through 2022, through the crypto nonsense, and the ramp-up of EVs. In large part, this was due to stable power growth, the replacement of coal with natgas, and the adoption of wind and solar. We were on track to go to 0-10% of historical emissions by 2040. https://www.c2es.org/wp-content/uploads/2022/11/2024-GHG-Tre...

LLMs changed that story; it is now a bad story. Emissions are back on the increase. Natural gas power plants are sold out for six years plus. We are on track to go back to 100% of historical emissions by 2040. EVs are a factor but were a factor in 2018-2022 as well. In terms of popular narratives, it's pretty accurate to say that LLMs singlehandedly reversed the only reason for climate optimism in the US.

[tptacek]

I'll read any source you have for this, but it seems unlikely given the low percentage of US energy consumption data centers account for. Your home air conditioning is most of American electrical consumption.

[shawndrost]

My statements about the recent history of declining US power sector emissions are pretty vanilla and I won't source them here, but they should be pretty easy to verify. What I'm asking you to take on faith (or to google) is that the positive trend was a mix of flat load growth and the energy transition (from coal to natgas, wind, and solar).

Load growth was flat 2005-2020, and now it's growing at circa 2%. Recent growth is almost all commercial and industrial, not residential [1]. There has been some manufacturing reshoring, etc, but the main driver is data centers [2].

"[D]ata centers consumed about 4.4% of total U.S. electricity in 2023 [and ~1.5% in 2014]." "[T]otal data center electricity usage climbed from 58 TWh in 2014 to 176 TWh in 2023 and estimates an increase between 325 to 580 TWh by 2028 [which would be circa 20% YOY growth]." [3]

Total US power use in 2023 was ~4000TWh. Compute power demand was at 4.4% of that and has been growing at ~20% YOY. (McKinsey forecasts 23% YOY growth and 2025 data center capex is growing 30%[4] but let's be conservative.) If 20% holds from 2023 through 2030 data center power demand will be at ~630TWh (16% of the size of the 2023 grid). If it holds through 2040 it will be at ~4000TWh (100% of the size of the 2023 grid). (No citations here, this paragraph is just analysis, and 630/4000 TWh are just 176*1.2^7/17. However these numbers are similar to the forecasts in [2] and [3].)

New data centers are more often powered by natural gas (and less by wind/solar) vs the grid at large, and will be true for the foreseeable future. I don't have a definitive citation for this but it's obvious to anyone close to this industry (as I am) and you can dig up any number of confirming citations about specific data center stories or utility IRPs (integrated resource plans). One concrete fact supporting this narrative is that gas turbines are sold out for the next ~6 years, which I don't believe has happened before in this century. Gas is on an absolute tear.

(Zero-carbon data centers that you read about, like Three Mile Island or next-gen geothermal, are specifically manufactured to disrupt the powered-by-gas narrative. They are about as common as data centers built in a way that revives coal plants; that is, they are a real but small phenomenon.)

So, that is the story. A) Power sector emissions were in decline, because of flat load growth and the replacement of coal with wind, solar, and natgas. B) This isn't true anymore; the grid is expanding quickly again, AI is the main culprit. C) At least for now, load growth is happening in a way that is dirtier than today's grid. C) If you believe that AI is going to continue to grow for 5-15 years like it is now, you also think that AI is (and especially, will be) a major driver of US greenhouse gas emissions.

And to add an anecdotal rebuttal to your earlier comment: sure, inference is just a few google searches worth of power. But now google searches have inference, so they are a few more times more power-hungry than they used to be. And when I code, I use way more inference than the number of times I used to google. And remember that when we say "inference", we are talking about last-gen LLMs; reasoning models (which are, as of a few months ago, the default in chatbots) are a chain of inferences which can be arbitrarily long. The current status quo is much more worrisome than your quip, and the problem compounds quickly if you extrapolate at all.

We have discovered a new and outrageously popular way to use compute, we built $455B worth of power-hungry data centers last year [4], and it is having (and will have) a huge effect on greenhouse gas emissions.

[1] https://www.eia.gov/todayinenergy/detail.php?id=65264 [2] https://www.utilitydive.com/news/load-growth-challenges-supp... [3] https://www.energy.gov/articles/doe-releases-new-report-eval... [4] https://www.ciodive.com/news/data-center-ai-cloud-infrastruc...

[tptacek]

Your source [2] provides numbers (and a graph) only for total US electrical usage (4.4 TWh or so). The number you cite for data center usage exceeds 2023 estimates for all data center usage globally. Do you have numbers for data center usage in the United States? It's the US TWh number that tells us whether data center usage is merely rising (of course it is!) or whether it's actually a significant component of all US usage.

This snagged you as well with your "recent growth" claim; I have no trouble believing it's true, but what you're saying and what I'm saying can be true at the same time.

I don't think there's anything productive I can do with your "$455B worth of data center" numbers, as there's no transform that takes me from buildout cost to electrical usage.

Respectfully: you said I was not just wrong, but "extremely wrong", so this should be easy for you to spell out. I appreciate the effort so far!

[shawndrost]

The below DOE link substantiates my quotes about US data centers and power usage, which I'll reproduce here. "[D]ata centers consumed about 4.4% of total U.S. electricity in 2023 [and ~1.5% in 2014]." "[T]otal [US] data center electricity usage climbed from 58 TWh in 2014 to 176 TWh in 2023 and estimates an increase between 325 to 580 TWh by 2028 [which would be circa 20% YOY growth]."

These quotes are ~compatible with your 4.4TWh number. If you still think the below DOE link is wrong and "...exceeds 2023 estimates for all data center usage globally" could you share why you believe that?

(Note that my "extremely wrong" is not directed at the literal text "LLM inference has a carbon impact of, like, a couple Google searches" but with the implication that LLMs have negligible carbon impact. If you think DCs were using 4.4% of US power in 2023 and growing at 20% YOY, and are a sizable-and-fast-growing carbon impact -- but that one LLM call is a small carbon impact -- I'll concede the latter and soften "extremely wrong" to "your original comment carried implications you didn't want".)

https://www.energy.gov/articles/doe-releases-new-report-eval...

[tptacek]

The 58->176 TWh interval from 2014 to 2013 clearly wasn't driven by LLMs; ChatGPT wasn't released until 2022. There were of course AI/ML models that preceded it, but nothing used at the scale LLMs are now. If your whole case is that technology writ large is driving data center expansion, that's fine; my argument is simply that it doesn't make sense to single out LLMs.

I think at this point though we understand the contours of our respective arguments! We don't have to keep litigating. Thanks for this!

[shawndrost]

My argument is really not about the 58->176 transition (which was slower than 20% YOY) but the rapid datacenter deployment that started around 2022. It is basically all LLMs (McKinsey says ~75% IIRC).

Anyway, yeah, thanks for the exchange!