If you work at a climate organization, someone on your staff has probably asked whether using AI betrays the mission. It's a fair question, asked in good faith.
We don't think it does. The reasoning has three parts:
The AI build-out is a real climate problem. Data center load is on course to more than double by 2030, and the new demand is being met with gas turbines and delayed coal retirements. The cost is not spread evenly: The turbines go up next to Boxtown in Memphis, where cancer risk already runs four times the national average.
Your own use is not what's driving it. A heavy user's entire year of prompting comes to roughly one round-trip flight from DC to San Francisco. Put an entire staff on AI text every workday, all year, and the total is about one home-month of electricity: under 0.001% of a typical organization's emissions.
The fight worth having is upstream. Whether a gigawatt of new demand gets met with solar and storage or with an on-site gas plant is decided in utility rate cases, permit hearings, and Clean Air Act enforcement. That's where the tonnage is. If you can find ways to use AI to advance those fights, you should.
The critics who call AI a climate disaster have the build-out right and the individual use wrong. The boosters waving per-prompt numbers around have the prompt math right and dodge the build-out. The third point is the reason we made the deck.
The whole deck is below if you'd rather just flip through it.
One: The build-out is enormous, and it's being powered with gas and delayed coal
Global data center electricity is on course to more than double, from about 415 TWh in 2024 to roughly 945 TWh by 2030, as much as all of Japan uses today. That's the IEA's Energy and AI projection.
The amount of capital going into this is absolutely mind-boggling. The four biggest U.S. hyperscalers have guided to roughly $725 billion of capital expenditure in 2026, up about 77% from $410 billion the year before. Goldman Sachs' baseline has annual AI capital expenditure reaching $1.6 trillion by 2031, and roughly $7.6 trillion spent cumulatively between now and then. Every one of the last four quarters revised these numbers upward. No amount of restraint at the prompt level is a counterweight to that.
The local story is worse than the aggregate one. In Memphis, xAI ran 35 gas turbines at its Colossus site without any permit at all through much of 2025, next to Boxtown, a neighborhood already carrying four times the national average cancer risk. Across the state line in Southaven, Mississippi, the company's second site could emit more than 1,700 tons of smog-forming NOx a year, by SELC, Earthjustice and the NAACP's calculation. The NAACP sued under the Clean Air Act in April.
At least 15 U.S. coal plant retirements have been delayed since January 2025, tied to data center demand and DOE emergency orders. Those 15 plants emitted nearly 1.5% of all U.S. energy-related CO2 in 2024.
And the companies making the loudest climate commitments are moving backwards. Google's emissions are up 81% since 2019. Microsoft's are running about 58% above its 2020 baseline.
Climate Power's Data Center Air Pollution Tracker scores the eight U.S. hyperscalers on two things: how much gas they burn on-site, and how dirty the grid is where they build. Built on Cleanview and EPA eGrid data, it's the clearest company-by-company view I've found.
Two: A year of heavy AI use costs about one plane ticket
Air travel turned out to be the right yardstick. Most people in this movement feel a little bad about flying and get on the plane anyway, having decided the trip was worth it. Using AI is the same kind of judgment call, made with much worse information.
A medium-heavy user, call it 35,000 messages a year, comes to under a tenth of one DC-to-SF round trip, as Google Flights counts it. Running Claude Code all day, every work day, comes to roughly one round trip. That second figure counts every call the agent makes rather than every instruction you type, and prices each one at 10 to 20 watt-hours. The IEA models an agentic request with reasoning at roughly 50 watt-hours of GPU electricity spread across four to six sequential calls, so mine sits in the same territory. Those are central estimates. Rerun them with every assumption tilted against AI and they rise to about one round trip and four.
To be candid, the level of uncertainty is lopsided between these two realms: The flight is known to within about 30%, while my lowest and highest AI estimates differ by a factor of about 50, depending on model, hardware, and grid. We know the flight well and the AI badly. The weakest inputs are my own — how many calls a day a heavy user really makes, and how much energy each one draws. Anthropic publishes neither.
So that gives you a test. If you've stopped flying altogether, then declining to use AI is consistent, and I respect it. If you flew to one conference last year, your entire year of AI use is a rounding error against that one trip. Using AI a lot is far less costly than flying a little.
Scale it to a whole staff and it stays small. Everyone on AI text, every workday, all year, comes out to roughly one home-month of electricity: under 0.001% of a typical organization's emissions, invisible against the travel, heating, and procurement that make up the real footprint. Video generation is the exception worth watching, because it costs far more per output than text does.
Two independent methods put the per-prompt figure in the same place. Google measured its own production systems in August 2025 and got 0.24 watt-hours for a median Gemini text prompt, about one second of a microwave running. Epoch AI, a nonprofit research institute, gets roughly 0.6 Wh for a 100-word prompt to a free chatbot, and that figure includes the model's share of its own training.
Which leaves the water. "A bottle of water for every email you write" is false. It traces to a single 2024 Washington Post graphic built on an assumption of 140 watt-hours per prompt, somewhere between 50 and 250 times too high. In May 2026, Shaolei Ren, the researcher whose figures that piece used, revised his own estimate down by about 33x, writing in an email quoted by Andy Masley that the 2024 number "should not be used to describe general AI/ChatGPT or today's optimized systems." Google measures the median text prompt at 0.26 mL, which is five drops.
Three: The emissions get decided in rate cases and permits, not in your chat window
Let's start on the core question: What if all this compute ran on clean power? That would solve it. Whether a gigawatt of new demand is met with solar and storage or with an on-site gas plant is settled by utility commissions, permitting agencies, grid operators and courts. The number of people at your organization using AI heavily has no impact.
Chatting with AI accounts for a tiny sliver of AI's energy usage. Ten billion AI text queries a day, at a generous 1 Wh each, works out to about 3.6 TWh a year. Set that against the 155 TWh that AI-specific data centers consumed in 2025 and all the world's chatbot text comes to something like 2% of AI data center power. Against all data center electricity, AI and non-AI together, it's under 1%.
Both inputs to that sum are deliberately padded high, so treat it as a ceiling rather than a measurement. It's back-of-envelope arithmetic. But the direction is not in question, and the other 98% is where everything actually happens: training runs, image and video generation, AI folded into search and social feeds and enterprise software, idle capacity. Total abstinence across your entire staff cannot move it.
Policy can. Utility tariff design decides whether data centers or ratepayers pay for the capacity they demand. Permits and Clean Air Act enforcement decide whether turbines like xAI's run. Dockets at FERC and the state public utility commissions (PUCs) decide who bears the marginal cost of new fossil load on a constrained grid. Disclosure rules decide whether any of us can see the picture at all, and right now only Google and Mistral have published credible per-prompt accounting.
This is also the part of our own work these tools have changed most. I recently mapped the entire Senate's positions on AI for an economic policy shop — voting history, coalitions, persuadable votes, allies, antagonists — in about fifteen minutes, work that used to take a week. The same method points straight at a PUC hearing: Which commissioners have moved before, and on what argument? The tracker I linked above runs on a national project database that would have needed a research department a decade ago. A few people keep it current now, against an industry with far more money.
The scale difference here is what I keep coming back to. Delaying a single 500 MW gas peaker keeps somewhere between 150,000 and 400,000 tons of CO2 out of the air each year, depending on how hard the plant would have run. A heavy AI user's entire year is about 600 kg, roughly that one DC-to-SF round-trip flight. The policy win is hundreds of thousands of times larger. Every hour spent auditing your staff's prompt count is an hour not spent in a rate case, and the exchange rate between those two activities is not close.
This is the personal carbon footprint all over again
There's a reason the individualizing frame feels so natural, and it isn't that it works.
The personal carbon footprint was popularized in 2004–2006 by BP, as Geoffrey Supran and Naomi Oreskes document, as part of a media campaign running over $100 million a year. It was built to move blame from corporate emissions onto individual choices, the same move Supran and Oreskes traced in ExxonMobil's advertising. Twenty years later most of us still use the frame instinctively, including inside organizations built to fight the industry that manufactured it.
The same frame is now pointed at AI. A movement that has spent decades on the receiving end of it should be quick to recognize it, and slow to take the bait.
What to do about it
Use AI internally, without pride or shame. The marginal climate cost of your own use is unmeasurable. Performing restraint that buys nothing is worse than neutral, because it costs you capacity you could be spending on the mission: winning a better climate.
If you manage people, be respectful and use a thoughtful change-management approach. The staff member who asked the question won't be talked out of it by a footnote. Say out loud where the organization has come down and why, what you're still uneasy about, and give people a sanctioned way to learn the tools together.
Carve out the harms you won't excuse. Environmental justice harms in communities near data centers, unpermitted turbines, runaway training and video compute. None of it gets a pass because the per-prompt math came out fine.
Fight for systemic change. Tariffs, permits, enforcement, disclosure. Intervene in the rate cases where your utility decides whether data centers or households pay for new capacity. Show up at the air permit hearings. Back the disclosure rules that would force companies to publish what their models actually consume. Advocate for regulating the AI companies for the public good. And use the tools on this work, not just on your newsletter: Fifteen minutes versus a week is the difference between launching a policy fight and skipping it. That's where the tonnage is, and where our actual power sits.
If your team is in the middle of this argument right now, we'd like to hear how it's going. Get in touch and tell us where you're stuck.