Corporate AI spending is hitting a wall, and the wall is the invoice. Microsoft canceled most internal Claude Code licenses inside its largest software division, Uber burned through its planned 2026 AI coding budget in roughly four months, and one company reportedly spent half a billion dollars in a single month after forgetting to cap employee usage. The Wall Street Journal called it rationing. That is about right.
The Microsoft tell
The detail that should worry every CFO funding an AI rollout: the tools are getting cut because people like them. Microsoft handed Claude Code to thousands of engineers, designers, and project managers in its Experiences and Devices group last December. By spring it had spread past engineering. Now it is being wound down, with a June 30 deadline, and staff are being pushed toward GitHub Copilot CLI.
The official internal reason was "toolchain unification," a phrase that means roughly nothing. The reporting around it, originally from The Verge's Tom Warren, points at the actual reason, which is that token-based pricing punishes you for adoption. Microsoft EVP Rajesh Jha framed the switch around being able to "shape directly with GitHub," which is the kind of thing an executive says when the real answer is the bill. The convenient detail is that June 30 is also the end of Microsoft's fiscal year.
There is a structural trap here worth sitting with. A normal software seat costs the same whether someone uses it for one hour or eight. A token-priced agent costs more the more useful it gets. So the better the tool performs, the faster the meter runs, and the harder it becomes to justify. Microsoft has its own models and its own coding assistant and it still got caught paying a rival to put a better product in front of its workforce.
Uber didn't have the cushion
Uber is the cleaner example because Uber cannot expense its way out of it the way Microsoft can. Per reporting citing The Information, CTO Praveen Neppalli Naga said the company spent its entire planned 2026 AI coding budget in about four months, with Claude Code usage across his roughly 5,000-engineer org climbing from 32 percent to 84 percent by March. Separately, Uber COO Andrew Macdonald told Business Insider that AI costs are getting "harder to justify."
That half-a-billion-in-a-month figure, by the way, comes from a single AI consultant talking to Axios about an unnamed client that never set usage limits. Treat it as a parable, not a data point. The mechanism it illustrates is real even if the number can't be checked.
So what's actually breaking
Ali Ansari, CEO of model-training firm Micro1, told Axios the enterprise is going through a "healthy swing" away from what he calls "tokenmaxxing," the habit of burning as many tokens as possible on the assumption that more usage equals more value. His sharper claim is the one buyers should chew on: "the reality of AI right now is that it only works for coding." If that is even half true, a lot of cross-department license spending is lighting money on fire to check the weather, which one CTO told Axios was an actual thing employees were doing with frontier models.
The uncomfortable counterpoint nobody at these companies wants on the record: when AI gets blamed for layoffs, the cuts may not be about productivity gains at all. CloudBees CEO Anuj Kapur put it bluntly to Axios, suggesting workforce cuts might be "the only lever they can pull" to offset the AI bills themselves. The tool you bought to save money becomes the reason you cut headcount to afford the tool.
Microsoft's Claude Code licenses end June 30. Watch whether other large enterprises follow with their own fiscal-year cleanups, or whether usage caps quietly become standard before the next budget cycle.
