One ChatGPT query is a second in the microwave

Nathan Labenz opens with a pocket calculator for a debate that has been running hot since late 2024. A single ChatGPT query, he says, uses roughly as much energy as running a microwave for one second. A cross-town car trip, a hamburger, and a hot shower each sit near 10,000 queries. Save one car ride with a chatbot and you have more than offset a year of ordinary use. A transcontinental flight lands around a million queries — more than a heavy user is likely to burn through in a lifetime.

The guest is Andy Masley, director of Effective Altruism DC and a former middle-school physics teacher who started writing because he wanted to win arguments at parties. He is a heavy user of current chatbots — Claude Opus 4.5 is "basically my new best friend" — and also wary of what more capable systems could do to society. That mix, he and Labenz agree, is closer to the EA median than the doomer-versus-accelerationist cartoon: excited about the tools, high on "P(weird)," and unconvinced by the specific philosophy arguments for a very high P(doom).

The mainstream story he wants to correct has two layers that people keep gluing together. Coverage of the data-center buildout is fair to treat as a huge industrial project with real uncertainty; what is not fair, he says, is assuming that more energy or water is automatically a catastrophe, or that your own prompting is a meaningful slice of your carbon or water footprint. He has even run into people at a large educational institution who wanted to block chatbot access for low-income students on environmental grounds — a stance that, followed consistently, would also make it hard to buy students books.

Part of the confusion is that digital goods feel weightless, so a big absolute number with no denominator sounds sinful. An old talking point still circulating is that a chatbot prompt uses ten times as much energy as a Google search; almost nobody adds that ten Google searches are still tiny. People also treat any *new* source of emissions as uniquely illegitimate, even though every car trip is new carbon that will linger for centuries. Data centers, Masley notes, are paradoxical buildings: they pile up millions of microscopic uses in one place, the way a warehouse of every toaster in America would look terrifying and still not tell you much about breakfast.

Masley's current party line, after adding inference, cooling, training, and the embodied carbon of the hardware, is about 0.3 grams of CO2 per median prompt — and he thinks that estimate runs high. That is roughly one hundred-thousandth of a typical person's daily emissions. You would need something like a thousand prompts to raise the day's footprint by 1%, and doing that would take on the order of ten hours of reading and typing. Ten hours spent on a computer is ten hours not driving, shopping, cooking, or sitting in front of a television.

If the query is that cheap, almost every offline activity you replace with it emits more. "Probably the single most climate-optimized way you can live your life is living in a big city, being vegan, and spending literally all your day on your computer," he says. He checks two of the three boxes. Comparisons to charging a phone (~20 watt-hours, about a third of a cent on a Detroit bill) used to be his favorite; he dropped them because most people have a three-rung mental scale — a little, a normal amount, a lot — and "phone charge" gets filed as normal. A watt is a rate; a watt-hour is energy. A median ChatGPT prompt, as best he can tell, is 0.3 to 0.6 watt-hours, or about a second on a 1,000-watt microwave.

Labenz stress-tests the claim from the other direction. A $20 ChatGPT subscription, if it were spent entirely on electricity at typical U.S. rates, would buy about 100 kilowatt-hours — roughly four days of a home's electrical use, or a 10% bump on a monthly bill. Electricity is only about a quarter to a third of U.S. carbon emissions, so even that cartoon upper bound is more like 3% of a household's total footprint, and the companies are obviously not turning the entire subscription into power. The interesting limit is economic, not thermodynamic: AI labs do not want to give away huge amounts of free energy.

The replacement effect is not hypothetical for Labenz. Over a month of his son's Burkitt lymphoma diagnosis, he ran ChatGPT, Gemini, and Claude in triplicate — a few hundred to maybe a thousand queries — to decide whether the family needed to move cities or rent a second house because of a possible mold problem. They stayed put and bought HEPA filters. Heating two houses for a winter, or relocating a family, swamps a lifetime of ordinary prompting. A friend of Masley's overheard two teenagers in a pizza shop: one mocked a classmate for using ChatGPT while calling herself an environmentalist, then ordered a meat-lover's pizza. "This kind of just sums up a lot of the debate."

The bill of materials is a trap. An eight-GPU H100 server node costs about $300,000 to buy and, over a four-year life, about $35,000 to power — only 10% of the purchase price. You might conclude that manufacturing is where the carbon lives. Masley says the price is mostly Nvidia's premium and a baroque supply chain, not a carbon invoice. An AI chip is designed to have electricity flowing through it until it wears out. Nobody worries about the embodied carbon of the wires in their walls; they worry about the electric bill.

Nvidia now publishes the carbon cost of its chips, and Masley's reading of the numbers is a ratio on the order of 20 to 1: the electricity used over the chip's life emits something like twenty times the carbon of making the chip. Labenz cross-checks with margins. If Nvidia is earning something like 80% gross margin, the cost of actually producing the node is closer to $60,000, which still includes TSMC, shipping, and everything else — a governor on how large the embodied energy can possibly be. Data centers also tend to sit on cheap, reliable power, which in the present grid often means dirtier-than-average electricity, pushing the operating-carbon share even higher.

When critics say he is hiding "embodied" costs, they rarely put a number on them. Poke at the hardware and it looks like 5–10% of the total. The line item that actually moves the estimate, Masley says, is training, a software cost rather than a factory cost. His current guess is that amortizing training roughly doubles the energy of a median prompt, from about 0.3 watt-hours to 0.6, with wide error bars. Even doubled, the personal-use conclusion does not flip. Worrying about the embodied carbon of an H100, he argues, is in the same family as worrying about the embodied carbon of household wiring.

A 20-mile sedan trip emits about 3 to 4 kilograms of CO2, on the order of 10,000 chatbot prompts. Avoid that one trip and you have 10,000 "free" queries. A transcontinental flight can exceed a million, sometimes two million — more than Masley, a self-described power user, expects to log in his life. Labenz re-derives the car number from a 20-gallon tank: about 100 pounds of gasoline becomes something like 300 pounds of exhaust after oxygen is added, or ~150,000 grams of CO2, versus ~0.3 grams per prompt. A full tank is hundreds of thousands of queries; one gallon of a 20-mile trip falls out near 10,000, within a factor of two of Masley's figure.

The water version of the panic is even more popular, and Masley wants it retired. A Washington Post piece treated writing a 100-word email as 10 to 20 queries, assumed chips had not gotten more efficient in five years, and folded in evaporative losses at hydroelectric dams. Stack those choices and you get "a bottle of water per prompt." Independent estimates, he says, put the real number around 2 milliliters — about 1/200th of the viral figure, or 200 prompts per bottle. An average digital alarm clock, off-site at the power plant, uses something like 3 liters of water a month, invisibly.

Three categories matter, and the coverage usually mashes them together. Consumptive use pulls water from a local source and evaporates or otherwise removes it — the kind that can stress an arid basin. Non-consumptive withdrawal takes water, often warms it, and puts it back; that is how a lot of thermal power plants cool. Pollution returns water dirtier than it arrived; data centers dose cooling loops with chemicals and are not zero on this, but Masley puts them far behind agriculture. A widely shared line that AI will use 50% as much water as the United Kingdom by 2027 is, on his breakdown, about 90% power-plant withdrawal that is returned, 5–7% consumed at the plant, and only ~3% inside the data centers. What sounds like half the UK flowing through server halls is more like 5% of UK *consumptive* use.

Inside the halls themselves, most of the water *is* consumed — evaporated — with maybe 20% returned, sometimes needing treatment. Data centers also prefer potable water so minerals do not gunk the equipment, which sounds like they are stealing drinking water; Masley says converting fresh water to potable is not the binding constraint in most places and can get cheaper with scale. For 2023, the last year with decent figures, he estimates U.S. AI's consumptive use at about 8 to 10 towns of 15,000 people. Spread across the country that is a planning problem, not a national water emergency. Irrigated alfalfa, much of it grown in deserts to feed animals, used something like a thousand times as much water as all of AI in 2023.

Zoom out from the prompt and Masley's confidence drops. Latest chips draw on the order of 1,000 watts, coincidentally about an average U.S. home. A 1-gigawatt campus is then a million chips and a million homes. Stargate's multi-year target of 5 gigawatts is 5 million homes, a metro on the scale of New York or Los Angeles. Electricity is only about a quarter of U.S. emissions, so a completed 5 GW Stargate is, on this back-of-envelope, around a 1% add to baseline U.S. emissions. Labenz summarizes the conversation so far as low-single-digit national emissions even if the mega-projects all get built — and water smaller still.

The buildings themselves still astonish him. "A single building could use like a percentage point of America's household electricity." At prompt scale he is not worried about electricity or water. At data-center scale, electricity is the challenge. Even aggressive growth looks to him like adding about 1% of America's irrigated corn on the water side — large, but a system that can adapt. Electricity as a share of the grid is orders of magnitude more serious, and that is where he can believe local environments get hurt.

Land is not the bottleneck people imagine. Labenz has 1 gigawatt from about 10 square miles of solar. The full Sam Altman-scale vision in the conversation — a $7 trillion buildout drawing something like 80 gigawatts — would then need about 800 square miles of panels, less than 1% of Michigan or of sunnier, emptier Nevada. From a cockpit, most of the country is empty. Data centers, Masley says, are "miracles of compactness," Moore's law applied to a building. Global energy use is on the order of 6,000 gigawatts; 80 GW is a bit more than 1%, and smaller than the increase expected from ordinary development in poorer countries over the same years.

Asked about Memphis and xAI's Colossus, Masley puts air pollution far above climate and far above water. Carbon is fungible: emit a ton here, cut ten tons there, and the atmosphere nets minus nine, because the harm is global and delayed. Air pollution is not. A coal plant with a pipe into your living room does not get to say the national inventory improved. Indoor air pollution already kills millions of people a year worldwide, more than WHO projections for climate deaths even into the 2040s and 2050s. U.S. estimates he has seen run from 30,000 to 100,000 deaths a year — in the same range as traffic, more than guns — with messy causality, but not a rounding error.

He does not want to over-claim on Memphis. Residents reported smelling gas; there were recordings of more turbines than permitted; later tests suggested it was not an ongoing problem. He would quadruple-check before litigating the specifics. The illustrative point survives: even a data center that "just uses the normal grid" is plugging into a system that is already hard on nearby, often poorer, neighborhoods. The Colossus area already had grade-F air. Forecasts have AI using a lot of coal and gas at least into 2030, alongside power-purchase agreements for new renewables whose economies of scale could even help the climate case. Water scare stories, in every locality he has checked, either fail the arithmetic or describe construction, not operations.

That ranking makes him impatient with the water discourse. If communities are going to fight a campus, he wants them fighting gas turbines and new coal, not cooling towers. Labenz's household takeaway is more pedestrian: upgrade the furnace filter — those things come out filthy after eight weeks — and treat a couple of HEPA units as a precaution, noise and all. Masley has not done the deep dive on indoor air for ordinary listeners, but he will say that the quality of the air where you live is probably a first-order health variable, and that data-center noise and land use need to be weighed against a large new taxable industry, not treated as automatic vetoes.

Electricity bills are the other local grievance. Lawrence Berkeley National Laboratory, Masley's favorite source on this beat, implies that almost all of the last five years' rise in *national average* electricity prices came from inflation and supply shocks — including a spell of Ukraine-war gas prices — not from demand. The statistic going around is a 35% rise in the average American bill since 2020; on his reading of that report, data centers contributed about zero of it. There *are* local utilities that have said they raised rates because of a data center. He has not yet found the water equivalent.

The next five years are different: U.S. electricity use is set to rise significantly for the first time since the financial crisis, with data centers as the main driver, plus some EV charging. The rise is still slower than the 1950–2000 trend. The new constraint is that we no longer want a fossil-fuel surge. A common mistake is to think higher regional demand locks in higher bills forever; if that were true, big cities would always be more expensive than the countryside. Economies of scale usually flatten the bump. The special risk with AI is that demand keeps stair-stepping up faster than the grid can catch it, so the "temporary" premium lasts for years.

His water take is the one he calls crazy. Phoenix-area deserts look like the last place you would put thirsty computers. Arizona's golf courses, statewide, appear to use on the order of 30 times as much water as all of the state's data centers, while the centers generate something like 50 times as much tax revenue per gallon. He would not add new water demand in Arizona. He would happily close the golf courses and stand up data centers: billions more for the state, no extra water. In much of the country the water problem is aging pipes, not a dry aquifer, and a large new customer can fund upgrades. The thumb-rule he *will* stand behind: do not site a campus that requires new coal, or keeping a coal plant open. Coal is uniquely bad for lungs and for climate.

He refuses a national siting formula. He grew up relatively comfortable in a left-behind area and would be "actively excited" about a medium-sized, taxable data center nearby; he would hesitate at a giant campus ringed with gas turbines. Race-to-the-bottom tax holidays are real, but the fact of the race is evidence that city governments see a prize. Labenz's policy moral is that the surplus exists to pay for federal-style air standards so jurisdictions cannot undercut one another on lungs. Masley wants targeted rules against significant air pollution that do not also block the long-distance transmission the green buildout needs. He keeps reminding listeners he is "just some guy" who went down a rabbit hole to win party arguments.

Labenz runs a hamburger through the same microwave conversion: a few kilowatt-hours of primary energy and a few kilograms of CO2-equivalent, or something like 5,000 to 10,000 ChatGPT queries. A hot shower comes back in the same band. Those are the personal "villains" that dwarf prompting, and even there Masley would rather talk about animal welfare than tofu-slab accounting. He loves current chatbots and fears near-term and long-term consequences. What he does not want is fellow AI critics bolting "and it uses a drop of water" onto every surveillance or misuse argument. Listeners notice when you reach for every tool.

On climate P(doom), he puts AI's contribution near zero and possibly negative. The International Energy Agency has projected that by 2035, AI applications might avoid something like four times the emissions that data centers cause — wildly uncertain, and mostly from smaller deep-learning systems, not the giant chatbots. Labenz notes that models finding new antibiotics, of the sort Jim Collins has described, can be millions of parameters on a couple of computers. Material-science gains on batteries and photovoltaics could dwarf data-center load. Masley's goofy calibration: running Minecraft in 1950 would have taken something like half the energy of the United States; now it runs on a laptop. Computing is so optimized that *what you do with it* swamps the joules.

The 80-gigawatt vision is still only a bit more than 1% on a ~6,000-gigawatt global energy system, and poorer countries getting richer will add more than that. Downside use-cases he takes seriously are not the data halls: recommendation apps that make shopping irresistible, manufacturing efficiency that raises total throughput without greening it, robotaxis that make a five-year-old's trip across town too easy and grow vehicle-miles. The climate only cares about tons in the air. If environmentalists show up, he wants them as soldiers for *how AI is used*, with data centers as a real but secondary front — not as the whole war.