Imagine your employer giving you a payment card reserved for artificial intelligence, with a monthly limit of $100,000.
According to the specialist outlet The Information, that was the limit available to some employees in Microsoft's cloud and AI division. That is no longer the case: it now stands at around $10,000 for most of them.
What is happening
These are limits, not actual spending: most engineers were nowhere near reaching them. But the signal is clear. Microsoft is also said to have cut by more than a third what it planned to spend internally on Anthropic's Claude models, a bill that was set to exceed a billion dollars a year. The company is pushing its engineers towards its own tool, GitHub Copilot, and towards OpenAI models. Claude, however, is still offered to Microsoft's customers.
At Meta, the number of employees using Anthropic's programming tool is said to have fallen from around 60,000 to 30,000, partly because of spring layoffs, but above all because the company is pushing its own tools.
And Uber had kicked things off in June: the company revealed it had used up its entire annual budget for AI programming tools in just four months. It has since capped spending at $1,500 per month, per employee and per tool.
Why the bill is soaring
Most professional AI tools are not paid for by subscription, but by consumption, in tokens. The more you use them, the more you pay, as with electricity.
Yet agents, those AIs that work alone on long tasks, consume enormous amounts. An engineer running several of them for hours can see their bill climb very quickly, without even realising it.
For months, many companies encouraged their teams to use AI as much as possible, so as not to fall behind. Then the bill arrived.
In American tech, a word has appeared to describe this behaviour: "tokenmaxxing", the practice of consuming as many tokens as possible, sometimes as a proof of modernity. At Uber, internal leaderboards had even encouraged employees to use AI as much as possible. This summer, a Microsoft executive, Jay Parikh, summed up the change in mindset in an internal message: the goal is not to consume less, but to get more impact per token.
What it changes
For companies, AI is moving from the status of experiment to that of a bill to be monitored, like electricity or travel. The question is no longer just "does it work?", but "is it worth what it costs?".
For employees, AI becomes a resource that has to be justified. Those who can get good results with little compute will have an advantage.
For AI makers, it is a warning. Their revenue often depends on a handful of very large customers, as Anthropic's IPO filing showed. When a giant tightens the screws, it shows up in the accounts. Hence the growing interest in much cheaper small models, capable of doing part of the work for a fraction of the price.
What we take away
The party of unlimited tokens is over. This is not the end of AI in business, far from it.
It is the beginning of its accounting. And as always, when a technology moves from enthusiasm to spreadsheet, it is the genuinely useful applications that survive.