Even Microsoft Can't Afford AI: The Token Trap That's Breaking Enterprise Budgets
Even Microsoft Can't Afford AI Anymore
Microsoft did something last week: starting June 30, it will cancel most internal engineers' Claude Code licenses and force a migration to its own GitHub Copilot CLI.
Six months ago, Microsoft was heavily promoting Claude Code, encouraging engineers to use it to reshape their workflows. Six months later, a single notice cut off access for everyone. According to The Verge (Microsoft cancels internal Claude Code licenses), engineers working on Windows, Microsoft 365, Teams, Outlook, and Surface are all affected. Nearly 100,000 people.
The reason is brutally straightforward: the bill is too expensive.
Microsoft has a market cap of $3.5 trillion. It has poured $13 billion into OpenAI. It has invested $5 billion in Anthropic. Anthropic has also committed to spending $30 billion on Azure computing power.
Yet it can't handle the Claude Code bill.
The problem lies in the billing model. Claude Code charges by Token — every word processed by the model costs money. When engineers use Claude Code to write code, a single request consumes dozens of times more Tokens than a normal chat. Code context, project structure, requirement descriptions, generated solutions, debugging suggestions — all of it is Tokens.
The more you use it, the more you pay. And the best thing about AI is precisely that it makes you use it more.
Uber is even more extreme. The CTO wrote in an internal memo: 95% of engineers use AI tools monthly, 84% have entered "agentic coding" mode, and 70% of online committed code originates from AI generation.
The numbers are beautiful. So is the cost — the entire AI budget for 2026 (about $3.4 billion) was burned through in four months.
Claude Code at Uber: after introducing it to 5,000 engineers, monthly usage soared to 85%-95%, with each engineer costing $500 to $2,000 per month in API costs. For a team of 100 people, just this one AI tool costs millions of dollars a year.
Uber's CTO summed up his situation in one sentence: "I'm back to the drawing board" — everything starts over.
For the past thirty years, software pricing has been "all-you-can-eat subscription" — pay a monthly fee, use as much as you want. Office 365, Jira, Slack — all follow this model. The more you use it, the lower the per-use cost, and the more the company profits.
AI has flipped this logic. The more you use it, the higher the per-use cost, and the more the company loses.
In traditional software development, the more engineers write, the more the company earns. In the age of AI agents, the more AI writes, the faster the bill to external suppliers grows. Every Token generated is consuming your cash flow in real time.
Bryan Catanzaro, Vice President of Applied Deep Learning at NVIDIA, recently publicly admitted: "For my team, computing costs are already far higher than employee salary costs."
AI was supposed to save you money. Instead, it spends your money faster than you could ever write code yourself.
There's an even more ironic detail.
Microsoft also built most of the computing infrastructure for Anthropic on Azure. But when its own engineers massively call Claude Code, Microsoft has to pay Anthropic per Token.
That's essentially funding its most direct competitor.
For Microsoft, GitHub Copilot is an "internal accounting cost" — resources circulate within Azure, with negligible marginal cost. Claude Code is a real "external bill." Even if the output is identical, the financial nature is worlds apart.
Cut the external bill, use the in-house tool. The functionality might be slightly worse, but the cost is controllable. Accounting logic beats engineers' technical taste.
However, there's another interpretation: Microsoft never really "couldn't afford it" — it "finished learning."
Let the competitor in as a sparring partner, expose the shortcomings of its own Copilot CLI, collect comparative feedback from engineers, iterate multiple times over six months. Once the gaps are mostly filled, pull the net.
This interpretation makes sense, but it only applies to Microsoft. Microsoft has cloud infrastructure, GitHub, and a large enough engineer base to serve as an "experiment sample." Most companies don't have these conditions — they aren't "learning then stopping," they simply "can't afford it."
AI tool pricing is shifting from "subscription plans" to "pay-as-you-go." GitHub fully switched to usage-based billing on June 1. Anthropic is switching enterprise renewals from seat-based to usage-based pricing. Claude Code's daily cost has doubled from $6 to $13. US AI software prices have risen 20% to 37% over the past year.
All signs point to the same outcome: AI will get more expensive, not cheaper.
Models are getting larger, Agents are multiplying, and call chains are getting longer. A single automated task might call multiple models in sequence, execute long-chain reasoning, consuming far more Tokens than today's chat-based interactions. Even if the price per Token drops, total consumption grows faster.
This is AI's pricing paradox: the more efficient it is, the higher the cost. The more indispensable it becomes, the more expensive it gets.