Is Worrying About AI's Power Bill Quietly Anti-Climate?
Description
Public anxiety about AI's electricity appetite is real and rising, but it is usually aimed at the wrong variable. The topline joules are large — the IEA sees data-centre demand roughly doubling to ~950 TWh by 2030 — yet the climate-relevant figure, carbon, stays modest (~1% of global CO2 by 2030, below aviation). Efficiency is improving ~40%/yr; Jevons rebound means total use still climbs, which only sharpens the real question: useful for what? On the evidence, a large slice of frontier compute is upstream of the climate solution itself — AlphaFold, GNoME's battery materials, learned fusion-plasma control, DOE's Genesis Mission, and LLMs compressing the grind of research. That same demand is also making firm clean power bankable (Three Mile Island restart, Meta's 6.6 GW nuclear, Google's geothermal/SMRs). The steelman survives, though: at the margin the present grid is dirty — PJM capacity costs spiked, coal retirements are being deferred, and ratepayers are absorbing billions — and most compute is not research. The episode steelmans the worry, then turns it: across every climate path the binding constraint is qualitative innovation, so an undifferentiated panic that chills the research-accelerating uses of AI is the one move reckless in all futures.
Sources & further reading (22)
- Energy demand from AI — IEAhttps://www.iea.org/reports/energy-and-ai/energy-demand-from-ai
- AI is set to drive surging electricity demand from data centres — IEAhttps://www.iea.org/news/ai-is-set-to-drive-surging-electricity-demand-from-data-centres-while-offering-the-potential-to-transform-how-the-energy-sector-works
- AI: Five charts that put data-centre energy use and emissions into context — Carbon Briefhttps://www.carbonbrief.org/ai-five-charts-that-put-data-centre-energy-use-and-emissions-into-context/
- How much energy does ChatGPT use? — Epoch AIhttps://epoch.ai/gradient-updates/how-much-energy-does-chatgpt-use
- How much electricity does AI consume? [2025] — Hannah Ritchiehttps://hannahritchie.substack.com/p/ai-electricity-2025
- AlphaFold: Five Years of Impact — Google DeepMindhttps://deepmind.google/blog/alphafold-five-years-of-impact/
- How AI Is Powering the Next Scientific Revolution — SK hynixhttps://news.skhynix.com/decoding-ai-how-ai-is-powering-the-next-scientific-revolution/
- Accelerating fusion science through learned plasma control — Google DeepMindhttps://deepmind.google/blog/accelerating-fusion-science-through-learned-plasma-control/
- Google DeepMind supports US DOE on Genesis — Google DeepMindhttps://deepmind.google/blog/google-deepmind-supports-us-department-of-energy-on-genesis/
- Applications of NLP and LLMs in materials discovery — npj Computational Materialshttps://www.nature.com/articles/s41524-025-01554-0
- Microsoft Nuclear PPA to Restart Three Mile Island — Data Center Frontierhttps://www.datacenterfrontier.com/energy/article/55142561/microsoft-nuclear-ppa-to-restart-three-mile-island-shows-hyperscalers-urgency-for-clean-energy
- Meta strikes 6.6 GW nuclear deal — Latitude Mediahttps://www.latitudemedia.com/news/meta-strikes-6-6-gw-nuclear-deal-to-fuel-its-ai-supercluster/
- Data center growth spurs PJM capacity prices by factor of 10 — IEEFAhttps://ieefa.org/resources/projected-data-center-growth-spurs-pjm-capacity-prices-factor-10
- PJM capacity prices hit record high — Utility Divehttps://www.utilitydive.com/news/pjm-interconnection-capacity-auction-data-center/808264/
- NY session wraps with data center moratorium — The Hillhttps://thehill.com/homenews/state-watch/5912788-new-york-data-center-moratorium/amp/
- Ill. Governor Pauses Data Center Incentives; NY Passes Moratorium — ENRhttps://www.enr.com/articles/63112-ill-governor-pauses-data-center-incentives-as-ny-lawmakers-pass-one-year-moratorium
- State Data Center Legislation in 2026 — MultiStatehttps://www.multistate.us/insider/2026/2/20/state-data-center-legislation-in-2026-tackles-energy-and-tax-issues
- AI and the environment — 2025 EPIC/AP-NORC poll, Univ. of Chicagohttps://climate.uchicago.edu/?p=12020
- The Age of AI: July 2025 — Verasighthttps://www.verasight.io/reports/ai-report-july25
- Making AI Less 'Thirsty' — Li, Yang, Islam, Ren (UC Riverside)https://arxiv.org/abs/2304.03271
- Meet the academics refusing to use generative AI — Naturehttps://www.nature.com/articles/d41586-026-00508-w
- AI and climate change — Energy and AI — IEAhttps://www.iea.org/reports/energy-and-ai/ai-and-climate-change
Script
Cold open
Here is a question that sounds almost backwards. What if worrying about how much electricity AI burns is itself the anti-climate position — not because the worry is selfish, but because it works against the very thing the worriers most want?
Frame
The numbers are real and rising. Every new data center is a headline; every query, a small charge of guilt. And the instinct is fair — we should ask what all this power is for. But most of the panic quietly assumes the answer is nothing, that the energy is waste. Full disclosure: this channel has argued, hard, that AI carries real and serious dangers — so this is not a defense of the technology in general, but a narrower claim about one specific worry. This episode takes that worry seriously, then asks whether, across every path off our climate trajectory, the energy is closer to an investment we cannot afford to skip.
How big is AI's power appetite, really?
Start with the honest, scary version. The International Energy Agency expects data centres to use about 485 terawatt-hours in twenty twenty-five, and to roughly double that — to around 950 — by twenty thirty. AI is the engine: demand from AI data centres is set to more than quadruple, the specialized servers growing about thirty percent a year. The United States and China account for nearly eighty percent of that growth. So it is a lot, it is fast, and it is concentrated — 950 terawatt-hours is more than Japan uses in a year. The topline is genuinely large.
Is that a climate problem, or just an electricity problem?
But electricity is not carbon — and the climate cares about carbon. The same IEA analysis puts data-centre emissions at roughly one percent of global carbon dioxide by twenty thirty — about 1.4 percent in the faster-growth case. Aviation runs around two-and-a-half percent, closer to three-and-a-half once you count its other effects. So the thing we are told to dread emits less than the flights we book without a second thought. The joules are large; the carbon is modest. Two different problems.
Doesn't efficiency just get eaten by more usage?
The sharper critics reach for the Jevons paradox: make something efficient and we just use more of it, so efficiency is a mirage. And the efficiency is real — AI hardware gets about forty percent more efficient each year, and Google cut the energy of a median prompt by a factor of THIRTY-THREE in one year. A typical text query now uses about zero-point-three watt-hours. Usage still climbs, because cheap, useful AI gets used more. But that is the part the Jevons argument skips: rebound is what the adoption of something useful looks like. Which leaves the real question — useful for what?
But does AI actually move the climate needle, or is that a vibe?
So, useful for what? Concretely. AlphaFold predicted the structure of over two hundred million proteins, released them free to more than two million researchers, and won the twenty twenty-four Nobel Prize in Chemistry. DeepMind GNoME proposed 2.2 million new crystal structures — including 52,000 candidate lithium-ion conductors, the raw material of better batteries — and labs have already synthesized hundreds. A learning agent held the plasma steady inside a live fusion reactor. The US Department of Energy now runs a research mission with DeepMind on fusion, materials, and earth science. Be precise about the strength here: the materials and protein results are proven, Nobel-recognized. The softer claim — that everyday chatbots speed up ordinary research — is real but more suggestive than settled. Grant only the hard cases, and the accelerant still points straight at energy, batteries, and materials: the climate bottleneck.
If demand is the villain, why is it reviving clean firm power?
And that demand is doing something climate pledges could not: making firm, clean power bankable again. Microsoft signed a twenty-year deal to restart a reactor at Three Mile Island — 835 megawatts of carbon-free power, a sixteen-billion-dollar commitment, back online around twenty twenty-eight. Meta contracted six-point-six gigawatts of nuclear. Google backed next-generation geothermal and small modular reactors. The AI buildout is not only consuming clean power — its appetite is helping pay to build it.
Does the climate math actually net out?
So does the climate math net out? Use the same scorekeeper. The IEA — whose scary projection opened this — also finds that widely adopting the AI tools we ALREADY have could cut about 1,400 megatonnes of carbon dioxide a year by twenty thirty-five. That is three to four times the total emissions of the data centres running them. But the IEA adds a warning in the same breath: right now, there is no momentum guaranteeing that adoption happens. Hold onto that — the payoff is real, and it is conditional on people actually using the thing.
So the worriers are simply wrong? Where are they right?
Here is where the critics have a real point — and they are not cranks: the degrowth and sufficiency camp, scholars like Kate Crawford who map the physical footprint of AI. The clean power is mostly future tense; the demand is present tense, and today grid is dirty. In PJM, the largest grid in America, the latest capacity auction cleared at a record 329 dollars per megawatt-day — data centers drove about sixty-three percent of that increase, roughly 9.3 billion dollars onto customers bills. In Maryland, coal plants slated to close are being kept open, with AI demand cited as the reason. And some of the harm is real and local — water drawn in already-stressed regions, ratepayers absorbing costs they never chose. So at the margin, right now, some of these joules really are coal, and someone else pays. That is the strongest version of the objection, and it is fair.
Forget whether the worry is valid — what is the worry doing?
Which raises the better question: not whether the worry is valid, but what it does. In twenty twenty-five, more than two hundred data-center bills landed in over forty states. New York passed a one-year moratorium. Ohio froze a tax program after its cost passed one-and-a-half billion dollars; Illinois paused incentives. At least eighteen states want data centers to pay their own way — much of which is just good policy. But notice the shape: an undifferentiated AI uses too much power, reaching for the brake without asking which uses it is braking. A moratorium cannot tell the model finding a battery electrolyte from the one writing spam. It slows both.
Forget the statehouse — what is the worry doing to the people who would actually build the way out?
The statehouses are not where this bites hardest, though — and I will be upfront that this next part is observation, not statistic. The University of Chicago found seventy-two percent of Americans worried about AI environmental impact, forty-one percent of them deeply — more than they worry about crypto, industrial meat, or flying, each of which has a larger footprint. The concern has drifted from the magnitudes. About thirty-nine percent of Americans now avoid these tools, and the environmental reason lands hardest on the young — the people with the most career ahead of them. Much of it traces to a figure — that one query drinks a bottle of water — that misreads the very research it cites: the engineer behind those estimates, Shaolei Ren at UC Riverside, measured water across a longer exchange, and says it depends entirely on local cooling and grid. The fear outran the footnote. And it is reaching high-leverage rooms: academics and writing faculty refusing the tools, some after reading about resource intensity. My read is plain: a quiet, moral opting-out, where one distorted figure talks capable people out of a tool that could multiply their work.
Turn
And that is the move to see clearly. It feels like caution — power is scarce, the grid is strained, so slow down. But caution about what? Across every serious climate future — optimistic, muddle-through, or grim adaptation — the binding constraint is the same: the speed of qualitative innovation. New chemistries, firmer clean power, better grids. That is the input every path runs through. And yes, most compute today is not science; it is ads, feeds, recommendation. But the answer to that is to STEER the energy toward the frontier, not to brake all of it blindly. A blanket panic that chills the research-accelerating uses of compute — lumping the fusion-control agent in with the ad-targeting model — is not cautious. It is the one move that is reckless in every future. Be clear about the bet, though: if the grid is still fossil-heavy in twenty thirty-five, or frontier AI stops paying off in energy and materials, this judgment flips. Done right, precaution protects the accelerant instead of throttling it. The work was never to shame the joules — it is to clean the marginal kilowatt-hour, and point the compute at the frontier instead of the feed.
Closer
So — back to the question. Is worrying about AI power bill anti-climate? Not the worry itself; the worry is useful. The anti-climate part is the worry that cannot tell consumption from investment — a joule that buys the exit from a joule that buys nothing. Efficiency matters. The dirty marginal grid matters. But next time someone says a data center uses as much power as a city, ask the question the panic skips. Power for what? On a trajectory that is failing by default, the riskier move may be to flinch at the tool that helps get us off it.