Microsoft Drops Claude Code After Burning Its Entire AI Budget

Microsoft is ending Claude Code licensing in its Experiences & Devices division by June 30, 2026. The official explanation points to toolchain unification. The real reason is more direct: thousands of engineers consumed the entire annual AI budget in a matter of months. Claude Code works too well for the finances to keep up.

Key Takeaways

• Microsoft terminates Claude Code licenses by June 30, 2026 in its Experiences & Devices division (Windows, Microsoft 365, Teams, Outlook, Surface)
• Token-based billing drove costs to $500–$2,000 per engineer per month
• At Uber, 5,000 engineers burned through the entire 2026 AI budget in four months

An Experiment That Cost More Than Anyone Planned

In December 2025, Microsoft distributed Claude Code licenses to thousands of engineers in its Experiences & Devices division, covering Windows, Microsoft 365, Teams, Outlook, and Surface. By spring 2026, the rollout had expanded to non-technical roles. In May 2026, Microsoft announced the program was ending on June 30. All affected teams are required to migrate to GitHub Copilot CLI. For context, see our earlier piece on Horizon: Claude Code and AI Agents Are Blowing Up Dev Budgets.

The official line is toolchain unification. The implicit signal is cleaner: Claude Code’s token-based pricing proved economically unsustainable at enterprise scale. Individual costs ranged from $500 to $2,000 per engineer per month depending on usage intensity, well beyond initial projections.

Uber’s experience illustrates the problem at scale. CTO Praveen Neppalli Naga said it directly: “I’m back to the drawing board because the budget I thought I would need is blown away already.” At Uber, approximately 5,000 engineers with access to Claude Code pushed adoption from 32% to 84%. The entire 2026 AI budget was consumed in four months.

These figures give a sense of the scale at Microsoft. If a single division generates that kind of consumption, a company-wide deployment would have represented hundreds of millions in unbudgeted annual expenditure. Pulling back isn’t a signal of distrust in Claude Code. It’s a signal about the limits of current pricing models. Others go further in restricting it, Alibaba banning Claude Code for its employees.

For additional context, the OpenClaw Framework, an open-source agentic system running on consumer Claude subscriptions at $200 per month, was consuming between $1,000 and $5,000 per day in API costs before Anthropic banned it from consumer plans. Intensive agent usage breaks pricing models designed for moderate use. On the infrastructure side, Anthropic also expanded Claude Code’s limits through the SpaceX deal.

The Broken Economics of Agentic Code

The core issue isn’t Claude Code. It’s the pricing structure of agentic AI systems. Unlike traditional software billed per license or per seat, LLMs bill per token consumed. An engineer using Claude Code heavily for complex reasoning tasks, parallel threads, and extended sessions generates exponential costs, not linear ones.

The Uber numbers are striking. Roughly 70% of committed code originated from AI, and around 10% of backend updates shipped with no human oversight. The tool is productive. Productive enough that engineers use it constantly. And that constant use breaks the budget math for procurement teams.

Bryan Catanzaro, Nvidia’s VP of deep learning research, put the structural shift plainly: compute costs now exceed employee costs for his team. An unprecedented shift in how a tech organization’s cost base is structured, one that forces a complete rethink of budget planning logic.

Gartner estimates that 25% of planned 2026 AI budgets will slip to 2027, with only 28% of AI infrastructure projects fully delivering on their business cases. Microsoft’s pullback isn’t an isolated incident. It’s the materialization of an economic model crisis affecting the entire sector.

As we analyzed in our coverage of Anthropic’s first profitable quarter, the company’s revenue is driven by growing enterprise demand. The paradox is worth noting: the more intensively Claude Code is used, the higher client costs climb, and the closer Anthropic gets to profitability. Both movements are happening simultaneously, in opposite directions depending on which side of the table you sit.

Also on Horizon:

• China cuts 12,000 university degrees made obsolete by AI
• 79% of parents already use AI to raise their kids
• GPT-5.5 Instant beats doctors on health answers

What This Signals for Enterprise AI

In the near term, other large organizations will make similar trade-offs. Procurement teams discovered in 2025-2026 that usage-based AI tools don’t fit into the annual budget planning models they’re built around. Cost unpredictability is incompatible with standard financial planning cycles at large organizations.

The market response will be a migration toward hybrid models. Unlimited per-seat access is giving way to capped, metered access billed like AWS: consumption tiers, spending alerts, and hard limits. Multiple AI providers are already moving in this direction to make their tools compatible with enterprise procurement processes.

For Anthropic, the question is whether this type of withdrawal creates a reputational risk or simply a contractual adjustment. The evidence points to the latter. Claude Code remains the de facto standard for AI coding, and Microsoft is migrating to GitHub Copilot CLI, not to a competitor. The internal alternative doesn’t solve the structural cost problem. The tool keeps gaining features, Claude Code adding artifact sharing.

In the medium term, the key question is what a viable economic model for agentic enterprise AI looks like. The tools that thrive will be those that enable fine-grained cost control without degrading the experience. This isn’t a question of model performance. It’s a question of commercial offer design.

Follow the story on Horizon.