This highlights what I dislike about the "nooo AI is actually not that bad for the environment" crowd: they wring their hands over specific quantities of water or watt-hours per query when it's indisputable that "AI", as an industry, is driving very bad environmental outcomes. We see it all the time, and here is another example. In their own words (as quoted by TFA):

[R]eaching our climate moonshot is getting harder. Growing our data center footprint to build out the infrastructure needed to make AI as helpful as possible to everyone requires energy and resources. [... O]ur AI infrastructure buildout is currently accelerating faster than the grid is decarbonizing.

We did not have to do this. We are breaking the world more and more for tools we did not need to build.

This is also not powered by green energy. There are gas pipelines being built directly to data centers for on site generation. Wait times for turbines have gone up years.

We are setting our future on fire so that we can accelerate enshittification.

I am very glad I don't have children.

Does it mean Google is currently constantly burning more than 5GW in average? This is insane. It’s like five regular 1GW nuclear power plants!

This blog post points to a really important discussion from the Internal Energy Agency (IEA) on the numbers that are given to estimate the cost of a single LLM query; the content is not so surprising, but seeing this in a credible report from an internal agency is important. I had trouble finding from the post (it is a screenshot of the text, and points to a tweet that points to the wrong document). It reasonates with discussions we have been having on Lobsters on whether the existing numbers given by AI companies are misleading, and whether it is reasonable to assume that the cost of model training are amortized by later usage (despite the explosion in number of new models being trained right now by many actors).

The quote is from the IEA's 2026 report, Key Questions on Energy and AI (see the PDF directly):

Box 2.1: How much compute does AI actually need?

The first official per-query energy disclosures from major AI companies paint a surprisingly modest picture. Google reports that the median Gemini text prompt consumes 0.24 Wh (Google, 2025b); OpenAI puts the average ChatGPT query at 0.34 Wh (Altman, 2025); and a bottom-up estimation by Microsoft Research arrives at a median of 0.34 Wh for frontier-scale models with over 200 billion parameters (Felipe Oviedo, 2025). These figures date from 2025 and are likely already lower. Google itself reported a thirty-three-fold reduction in energy per prompt over the 12 months to May 2025, driven by advances in model architecture and hardware. As inference optimisation and next-generation accelerators continue to improve, per-query consumption can be expected to fall further

Even taking the 2025 figures at face value and allowing for a generous blended average of 1 Wh to account for heavier workloads such as reasoning and multimodal generation, scaling to 10 billion queries per day would imply annual electricity consumption of roughly 3.6 TWh. To put this in perspective, 10 billion queries per day is comparable to the estimated volume of all daily Internet searches worldwide (Felipe Oviedo, 2025) and roughly four times ChatGPT's reported 2.5 billion daily prompts as of mid-2025 (TechCrunch, 2025). For context, this is less than 1% of the 485 TWh that data centres consume today and would represent only a small draw on the tens of gigawatts of new IT capacity that leading AI companies are individually seeking to secure.

This arithmetic raises a question that the available disclosures do not yet answer. If text- based inference at massive scale would account for such a thin slice of projected electricity demand, the bulk of planned capacity must be destined for other workloads, such as large-scale model training, video and image generation, autonomous agents running multi-step pipelines, or enterprise deployments not yet reflected in public usage figures. Yet none of the companies that have disclosed per-query metrics have offered a comparable breakdown of how their anticipated capacity will be allocated across these categories, nor how per-query consumption is expected to evolve as agentic and reasoning-intensive use cases become the norm. In practice, AI inference is diffusing rapidly across the full spectrum of digital services, from search and recommendation engines to customer support, public administration and autonomous multi-step agent workflows. Understanding the composition of future AI electricity demand, not only the cost of a single prompt today but the full mix of increasingly complex and embedded workloads driving tomorrow’s infrastructure build-out, would give energy planners and policy makers a much stronger basis for anticipating what lies ahead.

I just read that Google are planning to build a massive data center ~250km south of here. It is expected to use as much electricity as three Uppsalas, and that might double in the future. In addition to all the other issues, I think that might also make Torsboda a military target.

Google’s power consumption isn’t just growing – the rate at which it is growing is growing. We have a word for this: exponential growth.

No, such a function could just as easily be quadratic, or any other higher order of polynomial. An exponential function is a function whose rate of change is directly proportional to its value.

It's the paperclip maximizer in action!

At some level this is the same ancient story of unpriced externalities. Impose a real carbon tax and the market might be able to sort it out.

But in this case, the people participating in the market seem to have lost all pretense of rationality. The investors would just put the tax in a footnote and keep shoveling money. They're investing in PowerPoint slides about sending datacenters into space, for gosh sakes.

In the absence of a carbon tax, I hope local communities everywhere are taking advantage of this massive hype asymmetry to tax the heck out of any datacenters that get built.

NSA, GCHQ, ASD, CSE, and GCSB operate the most destructive category of compute infrastructure

The NSA Utah Data Center alone draws 65–90 MW continuously

Google fortunately is at least targetable though, no oversight on black budget FVEY datacenters.

[Comment removed by moderator pushcx: Don't pick fights.]

The Scaling Laws podcast published a good interview with Andy Masley on many concerns around data center growth, with electricity coming off as the best-justified criticism.

The main point made is that the IEA predicts world energy consumption to rise by 6000 TWh by 2030. About 3% of that rise is for data centres, including AI and cryptocurrencies.

https://www.worksinprogress.news/p/ai-is-bottlenecked-by-the-grid also claims that the bottleneck in the US isn't that we're not producing enough electricity, but that the queue to be connected to the grid is years long.