Microsoft’s discreet Claude Code retreat and the actual expenses associated with enterprise AI.

      In December of the previous year, Microsoft informed thousands of its engineers, product managers, and designers that they could utilize Claude Code, Anthropic's command-line coding agent, at the company's expense. By spring, this tool had extended well beyond the engineering sector and into various non-technical roles that would have typically waited years for such access during earlier phases of enterprise software. Within Microsoft, the deployment was presented as a learning initiative, while externally, the message was more straightforward: the world’s largest software company, with its own foundational models and coding assistant, opted to pay a competitor to provide a rival product to its employees.

      Now, six months later, this experiment is being curtailed. According to reports from Windows Central and others following the initial coverage by The Verge, Microsoft is canceling most direct licenses for Claude Code within its Experiences and Devices division—responsible for products like Windows, Microsoft 365, Outlook, Teams, and Surface. Affected engineers have been instructed to migrate to GitHub Copilot CLI by June 30, the last day of Microsoft's fiscal year. The official explanation is aimed at toolchain unification, but the underlying reason seems to be monetary.

      The reduction of Claude Code usage serves as the most convincing indication that, under current token pricing, the economics of enterprise AI coding are not sustainable. This isn't due to the tools being inadequate; in fact, they are proficient enough that engineers use them frequently, and such continuous use disrupts financial calculations.

      Uber sets a clear example of this issue. Praveen Neppalli Naga, Uber’s chief technology officer, disclosed to The Information in April that the company exhausted its entire AI coding budget planned for 2026 within just four months. By March, according to Naga’s statistics, Claude Code use had surged from 32 percent to 84 percent among his approximately 5,000 engineers, with individual engineers spending between $500 and $2,000 monthly on tokens. Now, around 70 percent of committed code at Uber is AI-generated, and nearly one in ten live backend updates is made by an agent without human intervention.

      “I’m back to the drawing board,” Naga commented, as the budget he anticipated was already surpassed. This statement succinctly encapsulates the situation: the projections were flawed because the variable in question, token consumption, does not behave like traditional user-based licenses and seats. In typical enterprise software agreements, costs are based on users. In contrast, token-based deals hinge on how much the model processes. Agentic coding requires extensive processing, leading to lengthy sessions and generating a substantial context, unlike the simpler autocomplete interactions that initially influenced pricing models.

      We have been observing this shift for months. Last November, GitHub halted new sign-ups for Copilot Pro and Pro+ due to the agentic workloads from paying customers causing expenses to exceed their monthly plan costs. The company acknowledged that cost structures designed for light assistance no longer applied.

      This situation is not limited to Uber or Microsoft; it reflects a broader industry phenomenon. Bryan Catanzaro, Nvidia's vice president of applied deep learning, stated to Axios in April that, for his team, the cost of computing now surpasses that of the employees who utilize it. Fortune reported in May that heavily utilized token-based AI tools can be more expensive per task than the human engineers they were supposed to support. A 2024 MIT analysis, widely circulated in finance circles since, indicates that AI automation currently appears cheaper than human labor for about 25 percent of roles initially thought to be replaceable.

      Contrasting this with spending projections, Gartner anticipates that global AI spending will reach $2.5 trillion this year, marking a 69 percent increase from 2025. The firm now categorizes generative AI in what it terms the trough of disillusionment, predicting that 25 percent of the intended AI budget for 2026 will be pushed to 2027 as many proofs of concept stall in procurement pipelines. An additional Gartner report from April revealed that only 28 percent of AI infrastructure projects fully meet their business objectives. This doesn’t reflect a technology maturing awkwardly; it indicates a market recalibrating its value.

      Microsoft's reevaluation aligns with this market adjustment and is not coincidental. There are two interpretations of their recent actions. The first aligns with Microsoft's official statement: that Copilot CLI is the strategic goal, ensuring engineers have ongoing access to Claude models through this avenue, and that the company desires a product it can directly mold with GitHub. This narrative is valid.

      However, it’s also a story Microsoft could have communicated any point in the past six months, but chose not to until now. The strategic rationale didn’t shift; what did change was the financial implications.

      The second interpretation is more challenging to dismiss. Microsoft is uniquely equipped to understand the true costs of enterprise-level Claude usage, given that its own engineers were among the heaviest users outside of Anthropic's clientele. Reports