When I scan the media looking for story ideas, one article on a subject doesn’t move me much. One story can be an outlier, a contrarian taking a swing, a writer talking their book. I file it and wait to see if anyone else picks up the thread. This past week, the thread turned into a rope. The Wall Street Journal, Axios and Fortune all landed on the same subject inside a few days of each other: The cost of AI has gotten so big, so fast, that corporations are pumping the brakes. They are rationing usage. They are pushing employees toward cheaper, weaker models. They are watching annual budgets vanish in three or four months. When there is that much smoke coming from that many chimneys, you stop debating whether there’s a fire. There’s a fire.
So let me say it plainly, because I don’t think it’s up for argument anymore. Companies are getting serious about putting AI on a fiscal leash. The all-you-can-eat phase is over, and the spreadsheet phase has begun.
The Numbers Stopped Being Abstract
For two years, the AI cost conversation was theoretical. Big numbers, far-off data centers, capex you could wave at without feeling. That’s done. The Journal’s Bradley Olson reported the figure that should stop every CFO cold: For companies using advanced AI coding tools, only 18% of what they spend on tokens turns into shipped product that reaches a real user. Read that again. Eighty-two cents of every dollar evaporates somewhere between the prompt and the customer. EntelligenceAI pulled that from more than 2,000 companies, so it isn’t a fluke.
The anecdotes are worse and funnier. Factory CEO Matan Grinberg told the Journal about an executive at a top financial institution whose employees were burning hundreds of thousands of dollars a month, some of them using premium-tier models to answer the simplest questions or just to make small talk. As the exec put it, if your daughter needs tutoring in algebra, you can probably find someone cheaper than Albert Einstein. Axios topped it: A consultant described a client who spent half a billion dollars in a single month after forgetting to put usage limits on employee licenses. Half a billion. In a month. Because nobody set a cap.
How did this happen? Because the model-makers built the trap and the customers walked in. Those early all-you-can-eat subscriptions were a subsidy, plain and simple. The labs lost money on power users and ate it to drive adoption. Employees, told to be AI-forward or be left behind, started tokenmaxxing, burning as much compute as possible to look like they were on the right side of history. Then the pricing flipped to usage-based, the meter started running for real, and the bill arrived. Uber blew through its entire 2026 budget for autonomous AI by March, per the Journal and Fortune both. Salesforce CEO Marc Benioff says his Anthropic tab will run about $300 million this year and openly wishes for a smart router that could stop sending small questions to expensive models. Meta CTO Andrew Bosworth put it best in an April memo to staff: all motion is not progress, and token usage alone is not a measure of impact of any kind
A Word on Microsoft, in Fairness
One detail deserves honest handling. The Journal reports that Microsoft, which cut Claude Code access for employees in several divisions, says the move wasn’t about cost at all. It was about standardizing the tools people use across the company. That account runs against how Axios and The Verge framed the same decision, and the duty here is to put Microsoft’s explanation on the record and take it seriously rather than wave it away.
But standardization and cost discipline are not opposites. They are usually the same move, wearing different clothes. Companies consolidate onto fewer approved tools precisely because sprawl is expensive, and reining in a dozen overlapping AI subscriptions is itself a way to keep the bills in line. Whichever word Microsoft prefers, it is walking in the same direction as everyone else in these stories. The denial and the trend can both be true at once.
So What Does This Do to the Curve?
This brings us to the consequences. Two big questions fall out of all this smoke, and they don’t have comfortable answers.
First, the adoption curve. For two years, we drew it as a straight line going up and to the right, the way every deck in every boardroom drew it. Fiscal discipline bends that line. Maybe it’s a healthy bend, the market finally separating real use from theater, which is roughly the case Fortune’s Jeremy Kahn makes when he argues that token spend was always a lousy proxy and the real value comes from reinventing how work gets done, not just automating the parts people dislike. Or maybe it’s the first genuine slowdown in a story that priced itself on never slowing down. I lean toward the former. Rationing is what grown-up technology adoption looks like. But anyone who tells you the curve is unaffected is selling something.
Second, and this is the one that keeps me up, what does rationing do to the AI labs and their trillion-dollar dreams? Anthropic just closed a $65 billion round at a reported $965 billion valuation. OpenAI is marching toward a listing. Those valuations were written during the all-you-can-eat era, on the assumption that usage only goes one way. Usage-based pricing solved the labs’ subsidy problem and handed their best customers a reason to cut back, in the same stroke. That’s a genuine tension, not a talking point. Now, the bulls have a real rebuttal, and the Journal gives it room: Corporate AI sales and usage have climbed faster than anyone forecast, Google now processes more than 3.2 quadrillion tokens a month, seven times last year, and plenty of smart investors call this the early innings. They may be right. Both things can be true. Demand can be exploding while individual buyers get disciplined about what they pay for. The labs are betting that the first force overwhelms the second. The customers in these three articles are the first data point, suggesting it might not be that simple.
The People Inside the Leash
Underneath the macro story are the workers. The same employees who were told a year ago to use everything are now being told to justify every token. The leaderboard that rewarded you for burning compute got taken down at Meta. And there’s a darker read, courtesy of CloudBees CEO Anuj Kapur, who told Axios that some layoffs may simply be the only lever companies can pull to offset their AI bills. Sit with that. The technology sold as the thing that would make workers superhuman is, in some shops, becoming the line item they get cut to pay for. That’s not the future anybody pitched.
The free-experimentation phase had a generosity to it. The spreadsheet phase will not. People will be asked to do more with weaker models, to think before they prompt, to treat a frontier model like the expensive specialist it actually is. That’s not the end of the world. It might even be good engineering discipline. But let’s not pretend it’s the same deal we were offered.
The leash is on. The only question left is how short they pull it, and how the companies that bet a trillion dollars on a longer leash plan to explain the difference.
