Microsoft just pulled the plug on most Claude Code licenses inside its Experiences and Devices division, effective June 30. The move comes barely six months after the company rolled out Anthropic’s agentic coding tool in December 2025, and it tells you everything you need to know about where enterprise AI spending is headed: straight into a wall.
The Numbers That Forced Microsoft’s Hand
Here is what happened inside one of the largest software organizations on the planet. Engineers using Claude Code were racking up $500 to $2,000 per month in API costs, each. For a division the size of Experiences and Devices, which builds everything from Windows to Surface to Microsoft 365 consumer apps, that math gets ugly fast. Multiply even the low end of that range across thousands of engineers and you are looking at millions per month in pure token consumption.
Microsoft’s solution is predictable but revealing: redirect engineers to GitHub Copilot CLI, the company’s own AI-assisted coding tool. That is not just a cost play. It is a strategic consolidation. Why fund a competitor’s model when you own the dominant code-completion platform and have billions invested in OpenAI? The surprise is not that Microsoft made this call. The surprise is that it took six months.
Uber Already Showed Us This Movie
If the Microsoft story sounds familiar, that is because Uber wrote the first draft. As Fortune reported in May, Uber burned through its entire 2026 AI tooling budget by April, just four months into the fiscal year. The company saw 84 to 95 percent adoption of AI coding tools by spring, which sounds like a success story until you realize adoption at that scale, at $500 to $2,000 per engineer per month, is a finance team’s nightmare.
Uber’s response was to impose a $1,500 monthly cap per tool per engineer, essentially rationing access to the tools its own developers had come to depend on. We covered the full breakdown of how Uber’s AI budget collapsed in four months and why the COO’s framing of it as a “good problem to have” does not hold up under scrutiny.
Microsoft is now the second major tech company to hit this exact wall. It will not be the last.
Consumption-Based Pricing Is the Structural Problem
The pattern emerging across enterprise AI adoption is not a bug. It is the business model working exactly as designed. Token-based pricing creates a paradox that no procurement team has solved: the more productive the tool, the more tokens it consumes, and the faster your budget evaporates. A senior engineer writing complex features with Claude Code can blow through $60 to $80 in API calls in a single deep coding session. Scale that across a workforce that is actually using the tool well and you get the budget craters that Uber and Microsoft are now staring at.
This is fundamentally different from traditional SaaS economics. A Jira license costs the same whether an engineer files two tickets a day or twenty. A GitHub seat does not get more expensive when a developer pushes more commits. But AI coding assistants operate on metered consumption, and the consumption curve is steep. Early pilots look affordable because adoption is low. By the time adoption hits 80 percent, the bill has grown by an order of magnitude, and finance teams are scrambling to explain a line item that did not exist 18 months ago.
Some enterprises have reportedly hit their annual AI tooling budgets in as little as three months. The industry is learning in real time that “unlimited access to AI” was never a sustainable procurement strategy.
What Comes Next for Enterprise AI Budgets
The Microsoft cancellation signals a broader reckoning. Companies will increasingly default to tools they own or control, which gives Microsoft (via Copilot), Google (via Gemini Code Assist), and Amazon (via CodeWhisperer) a structural advantage over pure-play AI vendors like Anthropic. If you are a CTO choosing between a $30-per-seat Copilot license with predictable costs and an API-metered tool that could run $2,000 per engineer in a productive month, the spreadsheet makes the decision for you.
Anthropic and other model providers will need to answer this with enterprise pricing tiers, committed-use discounts, or flat-rate plans that absorb the consumption risk. The current model, where the customer bears all the cost variability, is proving incompatible with how large organizations budget for developer tools.
The irony is sharp. AI coding tools genuinely make engineers more productive. The data from both Uber and Microsoft confirms that. But the pricing model punishes that productivity, creating a doom loop where success triggers cancellation. Until the industry figures out a pricing structure that aligns vendor incentives with customer adoption curves, expect more stories like this one, and expect them to come faster.