Software Forgot How to Be Expensive
Microsoft’s AI chief says the models cost too much. He’s right, and the reason runs through a 161-year-old observation about coal.
For a few years, my job was to mark a book of power and LNG to market. You pick up a particular kind of humility doing that. The number glowing on your screen is only ever on loan from the weather, the grid, and whoever is willing to sell you a megawatt-hour at two in the morning. Inputs have a will of their own, and they do not care about your forecast.
I figured I’d left all of that behind when I moved into software pricing, where the inputs hold still and the next unit is free. Then I read what the head of AI at Microsoft said this week, and I was right back on the desk.
Mustafa Suleyman, who runs Microsoft AI, told Bloomberg that the models his company has been buying, Anthropic’s among them, simply cost too much, and that he has his team building cheaper ones in-house. The line is buried in this week’s Tech In Depth newsletter, the sort of item that slides past between an earnings note and a product launch.
Sit with how strange that is. A software company worth north of three trillion dollars, and its AI chief is complaining about the price of his raw materials. Software bosses don’t usually sound like that. People who run refineries sound like that. So do airlines, and shipping lines, and anyone whose margins rise and fall on the cost of a physical thing. Software was supposed to be the way out of all that.
The honest version is this: AI is turning into a commodity business, and the people selling it have started to talk like the people who trade one.
The trick software pulled
For thirty years the whole magic of software lived in a single number, the marginal cost of the next copy. Write the program once and the ten-thousandth download costs you essentially nothing. That number is the engine under every fat software margin you have ever admired. It is why a company can hand its product to a million people for free and not flinch. “Free tier” was a strategy, not generosity, because the next unit really was free. The only question left was distribution.
AI does not get that number. Every answer a model gives has a power bill stapled to the back of it. Producing a token, roughly a word, means firing up a rack of very expensive chips that pull real electricity and throw off real heat, which then costs more electricity to haul away. The marginal cost of intelligence is not zero. It is electrons plus the wear on silicon, billed by the millisecond. For the first time in a generation, software is shipping a product whose next unit genuinely costs money.
That is what sits underneath Suleyman’s complaint. He is not annoyed at a supplier. He has worked out that he is in a business with a cost of goods sold, and nobody who grew up in software was trained for that.
The subsidy nobody puts on the invoice
What makes this so confusing is simple. The prices you see today are make-believe.
The big labs are selling AI for less than it costs them to make it. Sam Altman has said flatly that OpenAI loses money on its two-hundred-dollar-a-month ChatGPT Pro plan, the top tier, the one power users pay extra for. By one read of the company’s 2025 figures, it spent around $1.70 for every dollar it brought in, and it is expected to burn something close to $14 billion this year, with a few estimates running higher still. Anthropic has run a tighter ship. Axios reports its margins swung from roughly negative 94% in 2024 to positive 40% in 2025. Even so, its compute bill came in heavier than planned.
None of this is a secret, and none of it is stupid. It is a land grab, paid for by the deepest pool of capital the tech industry has ever assembled. Take the market now, fix the economics later. We have all seen this film. Remember when your Uber was mysteriously cheap, your burrito showed up for less than it cost to cook and drive over, your shipping was somehow always free? Someone else was quietly closing the gap. Cheap was a customer-acquisition expense wearing the costume of a fact of life. One by one, those companies had to start charging what the thing actually cost.
So there is a false floor under this entire market, a price that holds only because a handful of companies have agreed to eat the difference. Dario Amodei, who runs Anthropic, said it plainly at a conference in December: “There are some players who are YOLO, and I’m very concerned.” He was not doubting that AI matters. He was doubting the clock, the distance between being right about where this lands and being right about when the bills arrive.
Why “just make it cheaper” doesn’t save you
The obvious move is the one Suleyman is making. Drive the cost down. Build your own models, design your own chips, fit the model so tightly to the hardware that nothing leaks. And costs are falling, fast. By some estimates the price of getting late-2022, GPT-3.5-level performance has dropped something like 280-fold since then. Ramp’s data on what companies actually spend shows the average price of a million tokens sliding from about $10 to $2.50 in a single year. Per unit, intelligence is getting cheaper at a clip almost nothing else in the economy can match.
This is exactly where a 161-year-old idea wrecks the happy ending.
Get Abha Dalmia’s stories in your inbox
Join Medium for free to get updates from this writer.
In 1865 a young English economist named William Stanley Jevons was studying coal. The steam engine had just become dramatically more efficient, far less coal for the same work, so everyone assumed Britain would now burn less of it. Britain burned more. Cheaper steam made it worth dropping an engine into a thousand jobs that had never justified one before. Efficiency did not kill demand. It let demand off the leash. Coal use climbed.
Same story, different fuel. Economists call it the rebound effect. When each token gets cheaper, nobody quietly does the same work for less money. They do far more of it. They point models at problems that were never worth the trouble. The fashionable “agentic” setups, where a model plans, grabs tools, and grades its own homework, can chew through five to thirty times the tokens of a plain chatbot reply for one task. Reasoning models mutter to themselves for thousands of words before answering something you could have asked in ten. Price per token falls, tokens per task balloon, and the invoice goes up regardless.
You can watch it happen in the one figure that never flinches: electricity. Data-center power use has gone from roughly 58 terawatt-hours in 2014 to around 200 in 2025. In Ireland, data centers now draw close to a quarter of the country’s entire electricity supply. Out near the big American server farms, wholesale power prices have spiked, by as much as 267% over five years in some accounts, and ordinary people’s bills have climbed right alongside. The efficiency gains are real. Demand is eating them alive.
That is the flaw in the plan. Cheaper models will not shrink Microsoft’s AI bill, because cheaper models are precisely the thing that gets everyone to use more.
Welcome to heavy industry
So play it out. If your marginal cost is energy and silicon, and every price cut just grows your volume, the contest stops being about clever software and becomes something far older. It turns into a fight over who controls the inputs. Any commodities trader could have told you that. In the end it is the only fight there is.
Watch what the giants do, not what they say. Microsoft has set its 2026 capital-spending plan at around $190 billion. Its own finance chief pointed out that roughly $25 billion of that is just chips and memory costing more than they used to. Stack the four biggest hyperscalers together and they are on course to spend on the order of $725 billion in capital this year, up about three-quarters from last year, going by figures the Financial Times pulled from earnings reports. That is not a software budget. That is the capital weight of railroads and oil refineries. By some estimates these companies are now shoveling 45% or more of revenue straight back into physical plant.
And it is not only chips they are buying. They are buying power the way a smelter buys power, locking in dedicated, always-on generation through long contracts, more and more of it nuclear. Microsoft is bringing a reactor at Three Mile Island back to life to feed its data centers. Satya Nadella has taken to calling those data centers “AI factories.” Listen to that word. A factory is where raw inputs become product, and where the winner is whoever has the cheapest, steadiest supply of the stuff going in the door. As one industry write-up put it, the real edge in AI is now settling at the silicon and energy layers.
That is where Suleyman’s grumble is pointing. Custom chips “tuned to the harness,” he told Semafor. That is a sentence from someone running a production line, not someone shipping features.
If you’re buying, price for the real world
You are probably not in the business of building reactors. But if you have put AI anywhere near something that has to earn its keep, the lesson is the oldest one on any trading desk.
The cheap price is on loan. Build your product’s economics around something near the real cost of compute, not the subsidized sticker, and you will not be the company watching its margins evaporate the quarter the labs start charging in earnest.
Pick the model the job actually needs. Most production work calls for the frontier model about as often as a grocery run calls for a race car. Put the big model on the genuinely hard slice and a cheaper one on everything else. Skipping that step is not a kindness to your users. It is just being expensive for no reason.
Treat your tokens like the metered input they are. Give the spend an owner, a budget, a meter on the wall, the way a plant tracks its kilowatt-hours. Sprinkling AI everywhere because the demo felt like magic is how you wake up to a bill nobody can explain.
The people who come out of the next few years ahead are the ones already wired this way. They price against the input curve by reflex. They never confuse a subsidized number with a real one. They treat “free” as a tab somebody is running. For twenty years that instinct was a handicap in software, where the next unit really was free. It is about to be the most useful thing in the room.
The complaint is a coming-of-age
It is tempting to read Suleyman’s line as a hairline crack in the AI story. Even the believers think it costs too much. I read it the other way around. Industries only start complaining about input costs once they have grown up enough to have them. Nobody signs a power contract for a toy.
AI is leaving the stretch where the only question was whether the thing worked, and walking into the one where the questions are what it costs and who pays. That second stretch looks a lot less like software and a lot more like energy. Capital-heavy. Bound to its inputs. Won at the margin by whoever owns the physical layer. The subsidy keeping all of this out of view will not last forever. Public markets are coming for these labs, and public markets want to see margins.
When the real bill finally shows up on the invoice, the people sitting pretty will be the ones who already ran the numbers. They heard a software executive complain that his inputs cost too much, and they did not hear a man whining. They heard a forecast.
