Atoms, Electrons and the Token Economy

Every token is a few joules.

A language model is a machine for turning electricity into text. Each token it writes is the end of a chain that starts with a turbine spinning, a reactor splitting atoms or a panel catching light, and runs through transmission lines, transformers, power supplies and the multiply-and-add circuits of an accelerator.

Per answer the energy is small. Google reported in 2025 that a median text prompt to its Gemini assistant used about 0.24 watt-hours, counting the chips, the servers around them, idle capacity and cooling: less than nine seconds of television.[1] OpenAI’s chief executive put an average ChatGPT query at about 0.34. But a few watt-hours, multiplied by billions of prompts, agents that run all night and training runs that last months, adds up to the fastest growth in electricity demand that rich countries have seen in a generation.

A human brain does its thinking on about twenty watts. The machines we are building to think alongside it measure their appetite in gigawatts. This essay is about the layer in between: how electricity is made, how it is priced, how it reaches a data centre, and why the schedule of the grid has become the schedule of AI.

Flat for fifteen years, then a bend.

For most of the 2010s American electricity demand barely moved. Efficiency gains in lighting, motors and appliances cancelled out growth in everything else, and utilities planned accordingly: few new plants, slow transmission, a grid built for a flat line.

Data centres broke the line. Lawrence Berkeley National Laboratory estimates they used 176 terawatt-hours in 2023, about 4.4% of US electricity, and could use 325 to 580 terawatt-hours, between 6.7% and 12%, by 2028.[2] The International Energy Agency expects global data-centre consumption to roughly double, from about 415 terawatt-hours in 2024 to around 945 by 2030.[3] Five-year forecasts of US load growth have multiplied several times over in three years.[4]

The ranges are wide because nobody knows how fast AI will be adopted, how efficient the chips will get, or how many announced campuses will actually be built. What everybody agrees on is the direction.

The rack is the new unit of power.

A typical server rack in an enterprise data centre draws around nine kilowatts, about what seven American homes draw on average. An NVIDIA GB200 NVL72 rack, the workhorse of the current generation of AI campuses, draws roughly 120 kilowatts; its successor, the GB300, about 150.[5] The roadmap points past 600 kilowatts per rack within two years.

At those densities air cannot carry the heat away. The new campuses circulate liquid directly to the chips in closed loops, trading a little efficiency for far less water. The building becomes a heat exchanger wrapped around a power plant’s worth of load, and the most important number on the site plan is the size of the electrical connection.

A power plant is a heat engine with a schedule.

Most electricity in the United States still comes from burning something to boil water or spin a turbine. The key number for a thermal plant is its heat rate: how many British thermal units of fuel it burns to make one kilowatt-hour. A modern combined-cycle gas plant runs at about 6,400; an old peaker at 10,000 or more.

Multiply the heat rate by the fuel price and you have the plant’s fuel cost. At $3.45 per million BTU of gas, a combined-cycle plant burns about $22 of fuel per megawatt-hour; at $8, about $51.[6] That single line of arithmetic explains why gas prices move electricity prices, and why everyone building AI campuses watches the Henry Hub. Traders compress it into one number, the spark spread: the price of power minus the heat rate times the price of gas, the margin a plant earns for turning one into the other.[7]

Wind, solar and nuclear have almost no fuel cost, so they run whenever they can. Their economics live in the capital: what it costs to build them, and how many hours a year they produce.

The last plant needed sets the price for all.

Wholesale electricity in most of the US is sold in markets run by grid operators such as PJM, ERCOT and MISO. Every few minutes the operator stacks the available plants from cheapest to most expensive, the merit order, and dispatches them until supply meets demand. The marginal plant, the last one needed, sets the price that every running plant is paid.

This is why prices can sit near zero on a windy spring night and spike to thousands of dollars on a hot evening. In February 2021, during Winter Storm Uri, Texas prices sat at the $9,000-per-megawatt-hour cap for days.[8] Where transmission lines fill up, prices split by location. Each node on the network gets its own price, the locational marginal price, which is mathematically the shadow price of balancing supply and demand at that point. Analysts who trade power build production cost models that simulate this stack hour by hour, plant by plant, line by line, to forecast where each price will clear.[7]

Add a gigawatt of data-centre load that never switches off, and you move the whole stack to the right: more expensive plants run more hours, and the price rises for everyone on that grid.

Paying for the promise to be there.

Energy markets pay for electricity delivered. Capacity markets pay for something subtler: a promise that a plant will be available on the hottest afternoon three years from now. PJM, the largest grid operator in the US, runs an annual auction for that promise.

For years it was cheap. Then the data centres arrived. PJM’s capacity price went from $28.92 per megawatt-day for 2024/25 to $269.92 for 2025/26, then to the cap of $329.17 for 2026/27, $333.44 for 2027/28 and $325 for 2028/29.[9] Most of the jump was attributed to expected data-centre demand, and the last two auctions could not buy enough capacity at any allowed price: each came up more than 6,500 megawatts short of the reliability target. Those costs flow into every household bill in the region.

Power density is a fact of geography.

Vaclav Smil spent a career measuring energy in watts per square metre: how much power a source yields, or a use consumes, per unit of land.[10] In his words, “extraction of fossil fuels and thermal power generation proceed mostly with power densities of hundreds to thousands of Watts/m2,” while wind, solar and biofuels “produce energy at 0.1-20W/m2.” Cities, factories and high-rises consume at 10 to 9,000. Modern civilization, he argues, is a direct expression of the high power densities of fossil fuels.

A gigawatt campus fed from the grid needs a few hundred to a thousand acres. Fed by solar with enough overbuild and storage to run around the clock, it would need tens of thousands.[11] That is not an argument against solar; it is the reason power for AI is a question of transmission, siting and who lives near the wires.

Getting connected now takes longer than building.

To plug a new power plant or a large load into the grid, a developer joins an interconnection queue and waits while the grid operator studies what the connection would do to the network. At the end of 2025 about 2,061 gigawatts of projects sat in US queues, more than the 1,374 gigawatts of the entire installed fleet, and gas in the queue had nearly doubled in a year. Projects that reached operation in 2025 had waited a median of 61 months, against 22 in 2008.[12]

The equipment is queued too. Large power transformers take more than two years to deliver, most of them imported.[13] Heavy gas turbines are sold years ahead: GE Vernova’s backlog and slot reservations reached 116 gigawatts in mid-2026, with reservations now taken for 2031 delivery; Siemens Energy’s backlog is near 70.[14] In Texas, requests from large loads, mostly data centres, passed 400 gigawatts in mid-2026, while the load actually approved to energize was under 8.[15]

If the grid is slow, build beside it.

The fastest way around the queue is to not join it. Behind-the-meter projects put generation on the same site as the load, wired directly, so the campus needs little or nothing from the grid.

The examples are now familiar. Stargate’s first campus in Abilene pairs a 1.2-gigawatt grid connection with about 360 megawatts of on-site gas turbines.[16] xAI ran dozens of gas turbines in Memphis while it waited for grid power, and is now in court over air permits. Amazon’s plan to draw power directly from the Susquehanna nuclear plant was blocked by federal regulators in 2024, and was rebuilt in 2025 as a grid-connected contract for up to 1.92 gigawatts.[17] In Pennsylvania, Homer City, for half a century the state’s largest coal plant, is being rebuilt as a 4.5-gigawatt gas-fired campus with seven new turbines, for over $10 billion, with first power expected around the end of 2027.[18]

Behind the meter is faster. It is also a private grid, with its own emissions, its own reliability, and a bill that the public no longer shares, or no longer sees.

Firm, clean, fast: today, pick two.

An AI campus wants power that is firm (there around the clock), clean (for its owners’ climate pledges) and fast (this year, not 2031). No pathway delivers all three at once.

On-site combined-cycle gas is firm and relatively cheap, about $61 per megawatt-hour in one 2026 study, but it is not clean, and turbines are scarce. Solar with batteries is clean and quick to build, but making it firm raises the cost to about $130.[19] Restarting a retired nuclear plant is firm and clean; Three Mile Island’s Unit 1, renamed the Crane Clean Energy Center, is coming back, targeted for 2027, to serve Microsoft under a 20-year contract, but there are only a handful of such plants.[20] Small modular reactors promise all three, around 2030 at the earliest. Fuel cells arrive in months, at a high price, burning gas.

A little flexibility buys a lot of grid.

There is a cheaper way to find power: use the grid’s idle hours. Grids are built for the hottest hour of the year, so most of the time much of their capacity sits unused. A 2025 Duke University study estimated that the US grid could absorb about 76 gigawatts of new load if that load agreed to curtail 0.25% of the time, 98 gigawatts at 0.5%, and 126 gigawatts at 1%.[21]

Training runs and batch jobs can pause; some inference can move between regions. Google has signed demand-response agreements covering about a gigawatt of data-centre load, and field tests have cut a cluster’s power by a quarter for three hours without breaking its service commitments.[22] Flexibility is how a campus jumps the queue without building a private grid.

Underwrite time-to-power, not dollars per megawatt-hour.

Here is the counter-intuitive arithmetic. For a gigawatt AI campus, electricity is a small share of the total cost of ownership; one estimate puts energy at about $0.6 billion of $8.5 billion a year.[23] The chips dominate, and they lose value every month they sit idle.

So the cheapest power is rarely the best deal. Paying a clean-firm premium of $69 per megawatt-hour costs a gigawatt campus around half a billion dollars a year. A twelve-month delay in getting power can forgo several billion in revenue. The binding question for anyone financing AI infrastructure is not what the power costs, but when it arrives.

Efficiency will not save us from demand.

In 1865 William Stanley Jevons noticed that more efficient steam engines had not reduced Britain’s coal consumption. “It is wholly a confusion of ideas to suppose that the economical use of fuel is equivalent to a diminished consumption,” he wrote. “The very contrary is the truth.”[24]

The energy per AI answer is falling fast; Google reported a 33-fold drop per prompt in a single year.[1] Each fall has brought more use, not less: longer answers, agents that run for hours, models that think before they speak. Expect the same with electricity. Every efficiency gain in chips and models will be spent, and then some, on doing more.

Three horizons.

Forecasts disagree by a factor of several, so it helps to hold three horizons at once. The near one is visible in the queues: tens of gigawatts of AI load under construction, gated by turbines, transformers and interconnection. The AI 2027 scenario puts global AI power at about 38 gigawatts at the end of 2026, with 33 of them, 2.5% of US capacity, in the United States; its compute forecast reaches about 60 gigawatts by the end of 2027.[25]

The middle horizon, 2030, is where the bets are being placed: Leopold Aschenbrenner’s 2024 essay sketched training clusters of 100 gigawatts by the end of the decade, a fifth of current US generation.[26] The far horizon, 2040, depends on things we cannot yet price: small reactors at scale, fusion, compute in orbit, or a plateau in demand. The same team’s AI 2040 scenario sketches a world in which global energy use grows about eightfold, from about 19 terawatt-years in 2025 to a central 159 by 2040, with a range of 55 to 784.[27] The only robust strategy across all three horizons is to model the system openly and update as the queues move.

Civilizations are energy systems.

Vaclav Smil opens Energy and Civilization with a sentence that frames everything here: “Energy is the only universal currency: one of its many forms must be transformed to get anything done.”[28] The anthropologist Leslie White called it the basic law of cultural evolution: “culture evolves as the amount of energy harnessed per capita per year is increased, or as the efficiency of the instrumental means of putting the energy to work is increased.” Alfred Lotka went further in 1922: “natural selection tends to make the energy flux through the system a maximum, so far as compatible with the constraints to which the system is subject.”[29]

Intelligence is the newest thing we are converting energy into, and possibly the most consequential. The places that can build power will host the intelligence; the people who understand the energy system will understand the cost of thinking. That is why the next decade of AI will be decided in substations, turbine factories, permitting offices and interconnection queues, as much as in labs.