A word processor, once it exists, costs nothing extra to sell to the millionth customer that it didn’t already cost to sell to the hundredth. That is the secret behind a generation of software fortunes: the money goes into building the thing, while shipping another copy is essentially free. It is the opposite of manufacturing, where making one more unit eats up more than half of each dollar of revenue in raw materials and production.
An artificial intelligence (AI) chatbot behaves differently. It costs money every time it answers. With each character that appears on screen, a graphics processing unit (GPU) runs real calculations, and those calculations burn electricity and wear down expensive hardware. The industry calls this step inference. The “cost of making one more unit” that software treated as zero is not zero for AI.
The difference lands directly on the income statement. A mature software company typically posts gross margins of 70 to 85 percent; an AI service company runs closer to 50 to 60. OpenAI, the maker of ChatGPT, is estimated to have taken in about $3.7 billion in 2025 and still lost roughly $5 billion — it lost more than it earned. The culprit was neither research nor payroll, but the cost of inference itself: producing billions of answers a day. In truth, AI is a hybrid of software and manufacturing. The enormous up-front cost of training a new model resembles software development, except it recurs with every generation, and on top of it sits a variable cost attached to every answer. AI shoulders the most expensive half of each industry at the same time.
Stranger still, the unit price is collapsing while the bill keeps climbing. For a given level of capability, the cost of processing tokens has fallen to roughly one-fiftieth of its level three years ago. Yet total spending rises. As prices dropped, people started using AI far more — and in far more demanding ways. “Agent” workflows that reason in loops can burn dozens of times more tokens per request than a simple exchange. It echoes the nineteenth century, when more efficient use of coal led not to less consumption but to an explosion of it. Many analysts add that today’s low prices are a temporary floor, held below cost while companies fight for market share.
Follow this cost all the way down and what remains at the bottom is electricity. The International Energy Agency (IEA) projects that global data-center power consumption will more than double, from about 415 terawatt-hours in 2024 to roughly 945 by 2030 — close to 3 percent of all the electricity used on Earth. A rack of AI-specific GPUs can draw up to six times the power of an ordinary server rack. So data centers, like aluminum smelters before them, cluster wherever electricity is cheap. To the power grid, a data-center campus looks like a steel mill that appeared overnight: a new heavy-industrial customer, concentrated in one place and running around the clock.
The premise that built the digital economy — that copies are free — no longer holds for AI. AI is less like weightless software and more like heavy infrastructure, the kind that has to reckon with electricity, equipment, and location. That something is physically burning with every token is now attested by two ledgers at once: a company’s income statement, and a nation’s electricity-demand curve.