Microsoft just told us what token billing does to an enterprise budget

Microsoft is winding down internal Claude Code licences inside its Experiences and Devices division, six months after rolling them out, and routing future Claude consumption through Azure Foundry instead. The interesting fact is not that Microsoft chose its own toolchain — of course it did. It is that token-billed coding agents burned through the division's annual internal AI budget in a few months, and the company most aligned with usage-based AI in the world looked at the bill and said no.

The story is not the cancellation. Microsoft is, by every public account, still paying Anthropic. Engineers in Experiences and Devices, the group behind Windows, Microsoft 365, Outlook, Teams and Surface, have been told to migrate to GitHub Copilot CLI (the command-line version of GitHub's coding assistant) by 30 June 2026, which conveniently is the end of Microsoft's fiscal year. Claude remains available through Foundry, Microsoft's Azure-based AI governance and billing layer. The official framing is "toolchain unification". The unofficial framing, leaked through several outlets, is that the token bill was the trigger.¹

Per-seat and per-token are not the same shape. The dominant analytical frame for AI in 2024 and 2025 was SaaS displacement (the idea that AI tools eat into traditional per-user software budgets). It is a tidy story, and on the revenue side it is partly true — Claude Code is sold as a subscription, GitHub Copilot is sold as a subscription, and a lot of seat-licence budget has rotated into them. But underneath the subscription wrapper, the cost to serve a heavy user of an agentic coding tool is not a fixed monthly number. It is whatever the inference (the compute cost of running the model on each request) comes to.

That is the number that matters, and it is the number the SaaS-displacement frame quietly buries. A Salesforce seat costs Salesforce roughly the same whether the user logs in once a month or lives in the product. A Claude Code seat, or a Copilot seat used hard, does not. The unit economics on the buyer side are usage-shaped, even when the SKU is seat-shaped. At small scale this is invisible. At Microsoft scale, with thousands of engineers told to use the thing as much as they like, it is the budget.

This is what rationing looks like when it arrives. The standard reading of enterprise AI in 2025 was capability-bound: roll out more, deploy harder, the productivity gains will justify the spend. The Microsoft decision is the visible end of a curve that has been building for two quarters. Uber reportedly blew its 2026 AI budget on inference costs earlier this year. Procurement teams across financial services have started capping per-developer monthly token allowances. None of this means AI is not working. It means the cost is concrete and immediate, and the productivity gain is diffuse and slow to register in any financial statement a CFO can defend at the next board meeting.

The strongest counter-case here is that productivity measurement is genuinely lagged. Twelve months is not long enough to see whether a coding agent has compounded into faster shipping, better code, or lower attrition. The cost shows up on the invoice; the return shows up, if it shows up, in next year's velocity numbers and the year after's headcount plan. Calling the economics broken on a one-year window is premature. I take that seriously. But the point of rationing is not that the economics are proven broken. It is that the people writing the cheques have stopped being willing to wait and see.

Foundry is the part of this that should make Anthropic uncomfortable. When Microsoft routes its internal Claude consumption through Azure Foundry rather than a direct Anthropic subscription, three things change at once. Microsoft gets cost visibility and governance over how much each division spends. Microsoft inserts a billing layer between Anthropic and the end user. And Microsoft becomes the renewal counterparty rather than Anthropic.

Anthropic still gets paid. What Anthropic loses is the direct line to the engineer typing the prompt.

For Anthropic's revenue line this quarter, the change is probably neutral or positive — hyperscaler reseller agreements come with contractual minimums, and Foundry consolidates rather than reduces spend. For Anthropic's strategic position over three years, it is the structural risk that comes with selling through a hyperscaler that is also an investor and also a competitor at the application layer. The reseller becomes the customer. Usage signals, renewal leverage and pricing power migrate one layer up the stack. Anthropic knows this, which is part of why the recent Microsoft-Anthropic-Maia conversations matter — the question of whose silicon, whose billing layer and whose governance Anthropic's models flow through is not settled.

What I would watch. Three things, in order. First, whether other Microsoft divisions follow Experiences and Devices through the Foundry routing within the fiscal year — if they do, the "toolchain unification" framing is real and this is a Microsoft procurement story more than a Claude story. Second, whether enterprise coding-agent contracts start to include explicit per-developer token caps as a default term, the way mobile data plans did fifteen years ago — if they do, the SaaS-displacement frame quietly dies and a metered-utility frame replaces it. Third, whether Anthropic responds with pricing innovation that protects the direct relationship — committed-use discounts, enterprise flat-rate tiers, anything that makes going direct cheaper than going through a hyperscaler at scale.

The deeper point is the one I started with. The case for AI's economic impact has, for two years, leaned on the assumption that productivity gains would outrun inference costs at whatever deployment scale enterprises chose. Microsoft is the company that should be most able to absorb the wrong side of that assumption, and it has chosen not to. That does not mean the assumption is wrong. It means the assumption now has to be defended, in specific deployments, with specific numbers, against people who have seen one bill they did not like. The rest of the enterprise market has been watching Microsoft to learn what to do. It is now watching Microsoft learn what not to.

Footnotes

¶