Microsoft cut its engineers' AI budgets to $10,000 a month and your CFO is taking notes

Somebody at Microsoft once decided that an engineer in the cloud and AI group could spend up to $100,000 a month on AI models. That was the ceiling, according to reporting by The Information that other outlets have repeated. Most of those caps in the 60,000-person group are now dropping to about $10,000 a month, roughly $333 a day, and many queries are being steered to OpenAI models that Microsoft doesn't pay for separately. The company's forecast for internal spending on Anthropic, once at least $1 billion a year, has reportedly been cut by more than a third.

I expected this to be a story about rivalry. Microsoft sells Copilot, Meta builds MetaCode and Muse Code, and neither enjoys funding a competitor's revenue line. That is part of it. What caught me was a detail from Microsoft's Experiences & Devices organisation, which was told that most internal Claude Code licences would end by June 30, 2026, with engineers moved to GitHub Copilot CLI. People familiar with the decision said token-based usage costs had eaten through budgets far sooner than anyone projected, which makes this a finance problem first and a tooling preference second.

Meta's numbers make the same case from the other direction. Claude Code users there reportedly fell from about 60,000 earlier this year to about 30,000, partly through layoffs and mostly through a push towards its own tools. Meta still spent more than $105 million on Claude Code in a single 28-day period. Halving the number of people on a tool did nothing obvious to the bill, because the bill follows tokens, and I suspect the engineers who kept their access are the heaviest users.

When two companies with Meta's and Microsoft's money start policing per-engineer AI spend, I think every other operator should assume the same line item is heading their way. Gartner predicts AI coding costs will exceed the average developer's salary by 2028. Its analyst Nitish Tyagi argues that developers "optimize for speed and convenience over cost efficiency", and I don't hold that against them, since speed is what we hired them for. The problem is that nobody gave them a number to work within.

The survey data suggests most companies have already lost the thread. DoiT and Sapio Research asked 500 finance leaders and found that 79% had AI cost overruns in the past 12 months. A WitnessAI poll of 300 executives in July found 68% had at least some AI initiatives over budget.

The same DoiT data contains a fair objection to everything I am arguing. Companies with the most mature FinOps practices overran more often and by more: 89% of the very mature ones went over, by 30.9% on average, against 69% of early-stage ones, by 16.1%. Taken at face value, that says more budgeting discipline buys you bigger misses.

I read it differently. The companies with mature cost teams are the ones spending heavily enough to need them, and FinOps as most firms practise it reviews the cloud bill after the month has closed. Microsoft did something else in that cloud and AI group: it set a per-person ceiling in advance. It also did it bluntly, all at once, and engineers were reportedly unhappy about losing room to experiment with models. I'd call that the predictable result of engineering leaders leaving the number to finance until finance got nervous.

There is also the argument that both companies are curbing Claude for competitive reasons. Meta has issued guidelines limiting Claude and OpenAI's Codex for model-development work over fears of distillation, and Microsoft held back a newer Claude model while its lawyers reviewed Anthropic's data-retention terms. Both things are true, and neither applies to most readers of this magazine, who are not training a rival coding model but do receive a monthly invoice.

Anthropic, for its part, is doing very well. It told investors its annualised run rate hit $65 billion at the end of July, about seven times higher than a year earlier, and customer spending on its models through Microsoft's platforms keeps rising. Vendors will keep pricing these tools by the token, and the gap between what a coding assistant was forecast to cost per engineer and what it actually costs will land on somebody's desk.

Pull last month's model invoices this week and split them by team and by engineer. Give each team a monthly ceiling that its engineering lead owns, with a review when someone hits it, and have it in place before your next budget round. If you can't produce that split, your finance director will eventually pick one figure for everybody, and Microsoft has just shown you the ratio between the old cap and the new one: $100,000 down to $10,000.

Prompted by Meta and Microsoft Limit Employee Use of Claude AI Tools, RS Web Solutions (RSWEBSOLS).