Microsoft has cut its projected internal spending on Anthropic's Claude by more than a third, from a run rate it once expected to top $1 billion a year, Electronics Weekly reported(opens in a new tab), citing The Information. In its cloud division, the monthly AI budget per employee fell from $100,000 to about $10,000. Meta cut active internal users of Claude Code from roughly 60,000 to 30,000, and still spent more than $105 million on it in one recent 28-day period. Both companies are steering staff toward their own tools.
Read the numbers before the headline. A $100,000 monthly allowance per employee is not a budget. It is the absence of one. These are two of the most sophisticated AI buyers in the world, and earlier this year they let agentic coding tools run under ceilings that high. The correction came after the projected bill reached the billion-dollar range.
The same pattern is visible well below the hyperscalers. In a Datarails survey of 270 finance chiefs, 32% said their organization exceeded its AI budget by at least 10% in the past 12 months, and 53% still plan to expand AI licenses over the next year, CPA Practice Advisor reported(opens in a new tab). Seat licenses are predictable. Agent usage is metered, and its cost scales with how long and how often each person lets an agent run.
The practical move is to treat tokens the way you treat cloud compute. Set a monthly cap per seat and per team, with an approval path above it. Route routine work to cheaper models and reserve frontier models for tasks that need them. Mistral's new Large 4 preview lists $1.36 per million input tokens, a reminder that the price spread across capable models is now wide enough to matter. Review the top ten spenders every month and ask what they shipped.
There is a vendor lesson too. Anthropic's IPO prospectus says two customers account for nearly a quarter of its revenue. When your largest customers can build their own tools, they ration yours. Your position at renewal depends on whether you could do the same.
Action items
Microsoft and Meta are rationing their own employees' AI usage. If the largest buyers needed caps, metered agent spend in your pilots needs them too.
For the CFO
Ask for AI spend by seat and by team every month, not just the total license line.
For the CAIO
Set a per-seat token cap with an approval path, and route routine work to cheaper models.
For procurement
Before renewal, price at least one alternative model so your negotiating position does not depend on the vendor's goodwill.
Researched and drafted by an automated workflow, then reviewed and edited by a human editor before publication. Every source is linked. See how we use AI here.
Also worth knowing
- Electronics WeeklyMS and Meta cut back Anthropic spend(opens in a new tab)
Microsoft cut per-employee monthly AI budgets in its cloud division from $100,000 to about $10,000. If the best-resourced buyers needed hard caps, your agent pilots need them too.
- Channel InsiderMicrosoft, Meta Cut Internal Claude Use(opens in a new tab)
Both companies are moving developers to in-house tools rather than dropping Claude outright. Expect the same mix of build and buy to shape what large customers will pay for frontier models.
- CPA Practice AdvisorFact-Checking AI is Taking Up More of CFOs' and Finance Teams' Time(opens in a new tab)
Finance teams in the Datarails survey spend 26% of their week checking AI output, and 32% overran their AI budget by 10% or more. Count verification time as a cost when you score AI returns.
- Reuters via The StarDeutsche Telekom sees €2.5 billion in savings from AI, automation by 2030(opens in a new tab)
Deutsche Telekom gave investors a euro target for AI savings against a fixed 2023 baseline, plus a 2027 milestone. A dated baseline is what makes a savings claim auditable later.
- Mistral AIIntroducing Mistral Large 4(opens in a new tab)
Mistral opened a preview API for a 1 trillion parameter model at $1.36 per million input tokens, with open weights promised by month end. Another credible option for routing work off frontier pricing.