448 Terawatt-Hours — The Country That Is Not a Country

The Number I Printed and Then Could Not Put Away

In June of 2026 the United Nations University Institute for Water, Environment and Health — UNU-INWEH, Richmond Hill, Ontario, a name that sounds like a suburb and is a UN lab — published a report called Environmental Cost of AI's Energy Use. I printed the key points on one side because I am no longer a person who needs to pretend the paper is cheaper than the night.

The number that did the work was 448. Four hundred and forty-eight terawatt-hours. That is what the report said global data centers used in 2025. If you treat the campuses as a country, the country ranked eleventh in electricity consumption, roughly on par with France. France has people and trains and a nuclear argument. The campuses have a press release and a PUE.

By 2030, the same report said, the country could use 945 terawatt-hours — about three percent of projected world electricity, enough, if you like comparisons, to rank sixth. The International Energy Agency, which does not work for a suburb in Ontario, had already put a similar 945 in its base case. A later IEA update walked the near-term number up to about 485 terawatt-hours in 2025 and about 950 in 2030. Gartner, which sells the future in decks, said 448 in 2025 and 980 in 2030. I am not going to pretend these are the same thermometer. I am going to tell you they are the same weather.

Merritt asked which number to put in a footnote. I said put the range and the year and the author. He said a range is how amateurs look honest. I said a single number is how a slide looks like a fact. He wrote range and underlined it, which is how I knew he was tired.

What a Terawatt-Hour Is, If You Have Only Met a Kilowatt-Hour

Your house speaks kilowatt-hours. A kilowatt-hour is a toaster and an evening. A terawatt-hour is a billion kilowatt-hours. 448 of those is a toaster the size of a climate.

I sat in a capacity meeting, in a year that still had bagels, when a facilities person put a load curve on a slide and said we were "a small city." Someone from comms asked whether we should say city in external materials. Someone from legal said we should say campus. We said campus. We said efficient. We did not say France.

After the meeting I walked to a convenience store that still sold a paper map and I bought a bottle of water I did not need. I sat on a curb that had a stain and I did the arithmetic people do when they realize the city is not a metaphor. 448 TWh. The United States, the UN report said, already hosted nearly half the world's data centers. China about a quarter. Europe about fifteen percent, if you use the IEA's 2024 split. The rest of the planet got the queries and not the substations.

Inez does not speak terawatt-hours. Inez speaks the bill. The bill is a campus she does not enter.

Training Is a Bonfire. Inference Is the Winter.

The industry likes to talk about training because training is a headline. GPT-3, the UN authors wrote, took about 1.3 gigawatt-hours over thirty-four days. GPT-4, they estimated, took 50 to 70 gigawatt-hours over about a hundred days — forty to fifty-five times the earlier fire — and a carbon number around 25,000 tonnes, and a water number around 600 million liters. Those are estimates dressed as estimates. I am leaving them as estimates. The point is the scale, not a courtroom exhibit.

Training is a wedding. Inference is the marriage. The UN report, and every honest facilities person I have met, will tell you that once a model is out in the world the everyday questions are most of the electricity — on the order of eighty to ninety percent of the operational load. You do not need a wedding every day to empty a pantry. You need dinner.

I watched a product meeting decide that "usage would pay for itself in efficiency." Efficiency is real. Models get cheaper per token. Tokens get more numerous. Jevons is not a vibe. Jevons is a dinner that got inexpensive and then got ordered a billion times. I said so. The room did the laugh that means we will not put that in the notes.

The Meeting Where We Called It a City and Then Took It Back

I want to tell you about a specific afternoon, because the general afternoon is a pattern and the specific afternoon is a receipt.

A vice president asked for a one-pager that would make a board comfortable with a new region. The one-pager had a map and a green square and the word renewable. I asked where the megawatts would come from in year two, when the interconnect was still a queue. He said the utility had a plan. I said the utility's plan was a rate case. He said the rate case was not our slide. I wrote not our slide on a pad I later lost, which is how I keep the sentences I am not proud of.

That night I played golf because I did not know what else to do with a Thursday that had already been used. I shanked a drive on the third hole into a stand of trees that had never forgiven anyone. The man in the cart asked if I was all right. I said I had an electricity problem. He thought I meant the clubhouse. I let him think it.

208 Million Tons, Which Is a Country With a Name

The UN report's 2025 carbon number for the electricity behind the campuses was about 208 million tons of carbon dioxide — in the neighborhood of Argentina's annual emissions, if you like a flag. I do not love flag math. Flag math is how a slide becomes a sermon. I am putting Argentina in this paragraph once, because the AP did, and because a number without a neighbor is how people go back to sleep.

The same report said AI workloads were roughly a fifth of data-center electricity in 2025 and could be two-fifths by 2030. Two-fifths of 945 is a different country than the country we started with. The rest is video and storage and the ordinary internet you have already forgotten is a factory.

Merritt said I should not call 448 a country. He said it was a sector. A sector is a NAICS code. A country has a flag and a draft. I am calling it a country because the wattage already behaves like one and the vote has not caught up.

580 Billion, Which Is More Than the Oil

The IEA, in the 2025 World Energy Outlook season, put a sentence in an executive summary that I have watched people read twice. Investment in data centers was expected to reach about 580 billion dollars in 2025. Investment in oil supply was about 540 billion. The people who say data is the new oil now have a receipt that is rude even by their standards.

I am not going to tell you the 580 is all AI. It is halls and land and transformers and a lot of ordinary cloud. I am going to tell you that a sector that outspends new oil and still calls itself weightless is a sector that has hired a poet.

In the United States, the IEA said, data centers account for nearly half of electricity demand growth to 2030. By the end of the decade the country is set to use more electricity for campuses than for aluminum, steel, cement, chemicals, and the rest of the heavy list combined. That sentence is the one I wanted on the one-pager. It did not make the one-pager. The one-pager had a leaf.

The Aluminum Smelter Comparison, Which Is a Neighbor

AI-focused halls, the IEA said, can draw like a smelter and sit in a cluster like a religion. Nearly half of U.S. capacity is in five regional piles. A smelter at least has a union story people remember. A hall has a security guard who cannot say the name of the tenant. I stood outside a smelter once, in a life that was not this book, and the air had a taste. I stood outside a hall and the air had a sound. I know which one the brochure prefers.

A colleague who left for a utility told me the planners had started using large load the way doctors use mass. A mass is a thing you schedule. A large load is a thing you hope is real enough to build for and fake enough to not strand. The stranded cost is a kitchen. I said so. He said he knew. He was already tired.

Thursday, After the One-Pager

Inez does not come on Thursdays. I tell you that so you understand I was alone with the kitchen after I lost the pad that said not our slide.

I had the UN printout. I had a browser tab with the IEA summary. I had a bottle of water from the curb day, empty, which I had kept like a fool. I sat in the chair that faces the window and I did the arithmetic people do when they realize France is a competitor and not a vacation.

448. 945. 580. 540. The numbers are not a personality. They are a weather system. I thought about calling the vice president. I thought about calling the utility. I thought about calling a lawyer, who would have billed me in tenths and told me to document everything. I did not call them. I closed the tab. I left the printout on the sideboard, not under a bowl, on top, where Inez could see it and I could remember that the country had a body.

Inference at 4 p.m., Which Is a Peak

I want you to sit with a 4 p.m. in August. The kitchen runs an air conditioner. The campus runs a model that someone prompted from a phone that is also hot. The utility sees a hill. The hill used to be air conditioners. The hill is now air conditioners plus a straight line that does not go home. PJM prices that hill. Dominion socializes a piece of that hill. The prompt is free at the point of use. The hill is not.

A researcher told me, as if it were a comfort, that a single text query was "only a few watt-hours, maybe less if you're lucky." I said luck is not a tariff. I said a billion lucks is a plant. He said then people should prompt less. I said people should not be the fuse for a business model that sells unlimited. He went back to his benchmark. I went back to the printout.

The Last Week, the Power Version

There was not a single meeting. There was a week. The week had a launch. The launch had a blog post that said the new region was "powered by 100% renewable energy." I had seen the queue. I had seen the year-two megawatts. I had seen the diesel hours in a spreadsheet labeled contingency. I wrote a comment in a document that was too long. The comment said we should not say 100% if the 100% was a certificate. Someone from comms took the comment. The sentence stayed.

On the Thursday I boxed a plant I had never watered. I signed a paper that said I would not take things I had already put in my head. I shook a hand that did not know what to do with my face. I went to Merritt's and I watched him shred a pizza flyer. I told him I was done with the building. He said done is a feeling and a clause. I said I was still going to write. He said then write dull and specific and do not typeset the contingency file. I have not typeset the file. I have typeset the word. That is the line I can live with.

The Night We Did Not Stop the Run

I want to tell you about a training run that was not GPT-4 and was not famous and was still a weather event in a room with snacks.

It was a Saturday that had no business being a Saturday. The cluster was "warm," which is a word that means the cooling was doing math I did not understand and the facilities channel was doing English I did. A person I liked typed we can pause. A person I reported to typed we cannot. The cannot was a deadline. The deadline was a demo. The demo was a board. The board was a valuation. I have written that ladder before. I am writing it again because this time the ladder had a temperature alert.

I walked to the facilities cage, which is not a cage and is a room that smells like cardboard and hot dust. A technician had a laptop and a face that had decided not to be angry in front of me. He said the tower was "within envelope." I said the envelope was a warranty. He said the envelope was also a county, if the county ever asked. I said the county would not ask on a Saturday. He said that was the job.

We did not pause. The run finished. The demo happened. The board used a word like momentum. I used a word like envelope in a document that was later marked attorney-client, which is how a Saturday becomes a memory you are not supposed to have. I am having it anyway. The lake does not do attorney-client.

Merritt said I should not write the Saturday. He said a Saturday is a workplace and a workplace is a caption. I said the Saturday was a temperature. He said then write the temperature and not the slack. I have written the temperature. I have not written the slack. That is the line.

15 Percent a Year, Which Is a Pace

The IEA said data-center electricity had been growing about twelve percent a year and would grow about fifteen through 2030 — four times the rest of the world's electricity appetite. Fifteen percent is a number that sounds modest in a slide and is a doubling mood if you let it compound. I compounded it on a napkin, the other napkin, the one that was not wet-bulb. I threw that one away too. I remember the mood.

A planner at a utility, on a panel I should not have attended, said they were "reforecasting quarterly now." Quarterly is how you admit the last forecast was a children's drawing. I asked him whether the kitchen was in the reforecast. He said the kitchen was "residential class" and therefore a different model. I said the kitchen paid the model. He said that was a legal conclusion. I said it was a feeling. The moderator said we were out of time. We have been out of time since the queue.

France, Again, Because a Flag Helps a Kitchen

I said I would put France once. I am putting it twice, which is a broken promise and a method. 448 TWh is a country with trains. 945 TWh is a country with a seat at a table that does not exist. The table is the IEA's ranking. The seat is a joke until a minister treats it as a plan. I have watched American ministers treat campuses as a plan and African ministers treat blackouts as a weather. The UN report's whole moral, if it has one that is not a table, is that the plan and the weather are the same water cycle.

A journalist asked me, after I left, whether I thought data centers should be licensed like power plants. I said they already are power plants with better landscaping. She said that was a quote. I said it was not a quote. She used it anyway. I did not sue. I am not that kind of useful.

The Board Deck I Can Still See

Slide 1: a map with green squares. Slide 2: PUE trending down. Slide 3: a leaf. Slide 4: "community benefits." Slide 5: a queue position described as an asset. Slide 6, which legal took out: diesel hours. I have that deck in my head the way other people have a hymn. The hymn's chorus is efficient. The verse is a Saturday we did not pause.

Inez came the next Tuesday and asked why there were boxes. I said I had left a job. She said, "The model job." I said yes. She said good. She opened the window. She did not ask if I had a new job. I did not. I had a kiln in an LLC and a boredom and two notices. The notices became this tour.

Berkeley, December, a Range That Is a Mood

On December 20, 2024, the Department of Energy announced a report it had asked Lawrence Berkeley National Laboratory to write because Congress, in the Energy Act of 2020, had asked for an update to a 2016 study that had become a lullaby. The lullaby was: data-center electricity flattened. The update was: it did not stay flat.

I printed the DOE paragraph the way I later printed the UN paragraph. The furniture is simple enough to put in a kitchen. U.S. data centers used about 58 terawatt-hours in 2014. About 76 in 2018, 1.9 percent of the country's electricity. About 176 in 2023, 4.4 percent. The scenario band for 2028 was roughly 325 to 580 terawatt-hours — 6.7 to 12.0 percent of forecast U.S. electricity, if you believe the rest of the economy's denominator. The report said the 2014–2016 years had sat near 60. Then the installed base grew. Then GPU-accelerated servers stopped being a rounding error.

I sat with a person who still works in a lab that is not Berkeley and asked whether 176 was a scare. He said 176 was a history. He said 325 to 580 was a fork. I said a fork is a way to keep a minister from picking a number. He said a fork is how you stay honest when the shipment data is a weather. I said the kitchen does not receive a fork. He said the kitchen receives a rider. We were, for once, in the same paragraph.

The UN's 448 is a global 2025. Berkeley's 176 is a U.S. 2023. They are not a disagreement. They are two thermometers in two rooms. I have watched people collapse them into a fight because a fight is easier than two rooms. Keep them separate. The U.S. is about half the world's data-center electricity in the IEA's geography. Half of a country that is not a country is still a lot of kitchen.

A colleague forwarded the DOE release with the subject line finally. Finally is a feeling a person has when a number he already believed gets a federal stamp. I did not feel finally. I felt late. Late is the only honest feeling available once you have sat in a queue meeting and then read, a year later, that the queue was a national story.

The Plateau People Missed On Purpose

The 2014–2016 plateau is the sentence operators used to retire a conversation. We already did efficiency. PUE went down. The internet got bigger and the watts did not. It was true enough for a while, and then it was a costume. Efficiency is real. Rebound is also real. A better PUE on a denser rack is how you run twice the work and call the building green. I said so in a review. Someone wrote Jevons on a whiteboard and then erased it because Jevons is a name that makes a vice president look for the door.

I am not going to give you a lecture on a nineteenth-century economist. I am going to give you the cold aisle. We made the aisle more efficient. We put more heat in the aisle. The aisle did not write a thank-you note to the lake.

Gartner's Fivefold, Which Is a Vendor Weather

Gartner, in November 2025, put 448 terawatt-hours for 2025 and 980 for 2030, with AI-optimized servers going from 93 to 432. I am not a Gartner client. I am a person who has watched a vendor number become a board number because the board already wanted a steeper line. 980 is not a quarrel with 945. It is a different shop's weather. If your minister needs a higher number to approve a plant, he will find 980. If your minister needs a lower number to approve a campus, he will find a different slide. The kitchen finds the rider either way.

I asked Merritt whether I should pick a 2030. He said pick the range and stop decorating it. I am decorating it anyway, a little, because a range without a room is a PDF. The room is this: three serious shops — a UN lab, an energy agency, a research firm — all drew a line that triples a country. The argument is not whether the line exists. The argument is who drinks.

Inference, Which Is Most of the Country

The UN authors said AI was about a fifth of data-center electricity in 2025 and could be about two-fifths by 2030, and that inference — the answering, not the training — was on the order of 80 to 90 percent of the operational load once a model is loose in the world. I have used that split as a sedative. Training is a scandal you can date. Inference is a weather you cannot. GPT-4's 50 to 70 gigawatt-hours and 600 million liters, if you believe the estimate, is a training story. The chatbot in the toolbar is the rest of the country.

I sat in a product meeting where someone said we should "amortize training over a billion users." Amortize is a word that makes a lake look like a spreadsheet. I said the lake does not amortize. I was told I was being poetic. I was being a bill.

A researcher I like — I still like her — said the 80 to 90 was a climate, not a lab notebook. I said climates are what kitchens receive. She said then put it in the book as a climate. I am. If your minister is still giving speeches about a training run, your minister is one war behind. The war is the sidebar. The sidebar does not have a groundbreaking. The sidebar has a Saturday.

The $580 Billion Sentence, Which I Already Used and Am Using as a Room

I put the IEA's 2025 comparison in a paragraph and then walked away from it: data-center investment about 580 billion, oil-supply investment about 540. I am walking back because a sentence that rude deserves a chair. I watched a banker read it out loud in a room that had a view of a river the campus did not drink. He laughed. He said the punchline wrote itself. I said punchlines are how you retire a feeling. He said feelings do not close funds. I said kitchens do not close funds either, they just pay them. He ordered a sparkling water in a bottle and left it half full. The bottle was the meeting.

Oil is a hole and a ship and a war and a century of invoices. A data hall is a berm and a queue and a leaf. The IEA put them on the same ranking for a year and the ranking was not a metaphor. I am not going to do the cartoon where a GPU is a derrick. I am going to do the cartoon where a kitchen pays whichever ranking won.