The two biggest showcases of the coding agent boom found their own cap in the same month. Microsoft cut the monthly per-employee AI spending limit from 100 thousand to about 10 thousand dollars, and Meta reduced Claude Code users from about 60 thousand to about 30 thousand.
The numbers, with source
The numbers come from The Information reporting, without company confirmation. Microsoft expected to spend at least 1 billion dollars a year on Anthropic technology internally, and cut the projection by more than a third after telling the team to use GitHub Copilot, which runs OpenAI''s own models.
Meta migrated to its own tools, MetaCode and Muse Code, and still spent over 105 million dollars on Claude Code in a 28-day window.
The detail that matters to operators
The cap is a cap, not consumption. And customer-facing spend keeps rising, because Microsoft pays at least 2 billion dollars a year to Anthropic for Copilot features aimed at buyers.
Internal spend and product spend have different rules. The internal one was cut with a cap and routing; the product one keeps growing because it generates revenue. Mixing the two is what hides accounts that blow up.
The FinOps reading
The AI bill grows per user, and the corporate answer to the bill that blows up is not cutting people: it is setting a limit, measuring, and sending the work to where cost is lower. That is literally AI FinOps in operation, at the two largest technology companies in the world.
It is the same design we already documented here: an agent budget with kill switch to limit damage, and the 29% IaaS/PaaS waste rule to remind that the bill nobody measures is the one that grows.
Conclusion
If the two biggest showcases needed a cap, your operation needs one too. The difference is that theirs made the newspaper. A cap per user, measurement per tool and routing by cost are the three lines that separate an AI operation from an account without brakes.