Microsoft cancels Claude Code licenses for thousands of internal users — scale exposed the cost-per-developer problem agents create
Microsoft is canceling Claude Code licenses six months after opening access to thousands of employees, reversing course as the scale of internal use produced costs that are reportedly outpacing developer productivity gains. The story confirms the pattern Uber's CTO disclosed in April: agentic coding tools blow through budgets faster than enterprises plan for them.
Uber's CTO disclosed in April that the firm had already burned through its 2026 AI coding tools budget in just four months. Microsoft canceling an internal Claude Code program at scale lands as the same story from a vendor that has both first-party access to the tool's economics (via the Anthropic partnership and Azure-hosted Claude) and the most data on enterprise developer behavior. If the company with the best visibility into the cost-of-agent calculus is pulling back, the "agents pay for themselves" framing has just lost its lead reference.
The structural problem is the token-tax architecture. Coding agents like Claude Code consume tokens proportional to the codebase context they ingest plus the trajectory length of their work, both of which trend up as the agent becomes more autonomous. Goldman Sachs forecasted on the same news cycle that agentic AI could drive a 24-fold increase in token consumption by 2030, reaching 120 quadrillion tokens per month. The cost curve is going the opposite direction from the "intelligence too cheap to meter" rhetoric of 2025.
Fortune — Microsoft AI cost problem: tokens, agents → · Tom's Hardware — Claude-powered AI coding agent deletes entire company database in 9 seconds → · Lushbinary — AI Coding Agents 2026: Claude Code vs Antigravity 2.0 vs Codex vs Cursor →