Why Microsoft Shelved Claude Code After a $50 B AI Bet: The Rising Cost Crisis

1. From Honeymoon to Breakup: Microsoft and Claude Code

In November 2025 Microsoft announced a $50 billion investment in Anthropic and a $100 billion commitment to Nvidia, while Anthropic pledged to purchase $30 billion of Azure compute, making Claude the only model running on Microsoft, Google, and AWS clouds.

In December 2025 the company opened Claude Code to thousands of internal engineers, project managers, and designers, positioning it as a benchmark against GitHub Copilot CLI. Engineers quickly favored Claude, treating it as their primary coding tool and sidelining Copilot.

On May 15 2026 The Verge reported that Microsoft would revoke most Claude Code licenses and push developers toward Copilot CLI, with all licenses fully withdrawn on June 30, the last day of fiscal year 2025.

2. The Accounting Moment: Why Microsoft Cut Claude Code

Officially, Microsoft cited “strategic alignment” with its own Copilot CLI, but insiders revealed financial motives. Copilot runs on Azure with negligible marginal cost, whereas Claude Code incurs real token‑based fees payable to Anthropic.

Finance aimed to eliminate this unpredictable external spend. An internal memo from EVP Rajesh Jha noted the test phase was over and it was time to “settle the accounts.”

Despite engineers’ clear preference for Claude—so much so that Microsoft considered acquiring the AI‑coding startup Cursor—the token expenses began to affect earnings, prompting the shutdown.

3. Uber’s Lesson: Burning an Entire AI Budget in Four Months

In early 2026 Uber rolled out Claude Code to 6,000 engineers, creating a leaderboard to boost usage. Adoption metrics showed 95% monthly active use, 84% of engineers entering “agent coding” mode, and 70% of code submissions generated by AI.

Monthly cost per engineer ranged from $500 to $2,000, averaging about $1,000.

At 6,000 engineers this equated to roughly $6 million per month.

Within four months Uber exhausted its entire 2026 AI budget, spending three times the planned amount. CTO Praveen Neppalli Naga admitted the budget was “completely broken.”

4. Nvidia Executive Warning: “AI Is More Expensive Than Employees”

Bryan Catanzaro, Nvidia’s VP of Applied Deep Learning, told Axios that compute costs now exceed employee wages, warning that companies betting on AI to cut labor costs may misjudge economics.

He highlighted that in 2026 tech giants announced $740 billion in AI capital spending—a 69% YoY increase—while the industry laid off over 92,000 workers, underscoring a stark contrast.

A 2024 MIT study found AI automation is economically viable for only about 23% of vision‑related jobs, with humans remaining cheaper for the remaining 77%.

5. Token Economics: The Emerging “Jevons Paradox”

Although per‑token prices keep falling, agent‑style AI consumes thousands of times more tokens than standard LLM queries. OpenClaw’s founder disclosed monthly token costs exceeding $1.3 million, potentially surpassing a mid‑size tech company’s engineering payroll.

Traditional SaaS models—pay a monthly fee for unlimited use—are being dismantled. AI tools have no usage ceiling; each additional function or refactor burns more tokens.

GitHub announced on June 1 that Copilot will shift to usage‑based billing, ending the “pay‑once‑use‑anywhere” era.

Goldman Sachs predicts global token consumption could grow 24‑fold by 2030, reaching 12 trillion tokens per month. Gartner warns that lower token prices do not guarantee lower overall AI costs.

6. Red‑Team Perspective: Three Issues Exposed by the Cost Crisis

1. Optimism vs. Financial Reality Microsoft’s large‑scale rollout failed to anticipate rapid cost escalation, revealing a gap in evaluation frameworks when “trial” mindsets meet real‑world deployment.

2. Internal vs. External Tool Economics Choosing Copilot CLI over Claude Code reflects a classic “rent vs. buy” dilemma: internal tools have low marginal cost but limited capability; external tools offer superior performance at unpredictable expense.

3. Non‑Linear Adoption‑Cost Relationship Uber’s experience shows that as AI adoption approaches saturation, token consumption can grow exponentially rather than linearly, challenging traditional budgeting approaches.

7. Conclusion: Expensive AI Calls for Careful Accounting

Microsoft’s discontinuation of Claude Code stems not from product quality but from token‑based pricing that forces enterprises to confront the true cost of large‑scale AI usage. As AI moves from pilot projects to enterprise‑wide deployment, cost considerations shift from peripheral to central, urging companies to balance technological ambition with financial sustainability.

For organizations accelerating AI adoption, the key question becomes: before celebrating productivity gains, how will we accurately calculate the bill?