Microsoft, Uber Reel as Agentic Coding Costs Balloon

Microsoft has canceled most internal licenses for Anthropic’s Claude Code and Uber reports its 2026 AI budget was used up in four months after a rapid rollout of agentic coding tools.

Microsoft began winding down access in mid-May, with most licenses in its Experiences and Devices division scheduled to end on June 30. Engineers in that group had widely adopted the tool, and an internal workplace report cited productivity gains of up to 80 percent. At deployment scale, however, token-based billing produced costs the company deemed unsustainable.

Uber deployed Claude Code to about 5,000 engineers and exhausted its full 2026 AI allocation by April, according to the company. Praveen Neppalli Naga, Uber’s chief technology officer, told reporters that the budget was consumed faster than expected and quoted a return to planning: “I’m back to the drawing board, because the budget I thought I would need is blown away already.” Internal estimates put per-engineer bills between $500 and $2,000 per month, and some teams report roughly 70 percent of committed code originates from AI tools.

Agentic coding tools combine large language models with execution environments and orchestration layers to write, run and validate code with limited human direction. Their billing typically charges for units of text processed or generated, known as tokens. Automated testing, repeated runs and multi-step planning multiply token use, driving consumption beyond per-seat forecasts.

Industry data point to broader budget pressure. A 2025 survey by Mavvrik found 85 percent of companies missed AI cost forecasts by more than 10 percent, and 84 percent said AI spending reduced gross margins by over six percentage points. Big tech capital expenditure on AI reached about $650 billion in the first quarter of 2026, and the share of teams focused on AI financial operations rose from 31 percent to 63 percent within a year.

Anthropic projects a $10.9 billion second-quarter revenue forecast, reflecting strong enterprise demand for its products. At the same time, multiple companies are adding financial controls to deployments that expanded in late 2025. Measures in use include quotas, default routing to smaller or specialized models, result caching, token budgets per engineer or project, internal usage leaderboards and approval gates for high-cost automated runs.

Finance and engineering teams are testing governance and routing strategies intended to limit expensive calls to large general-purpose models while preserving developer access to automation. Firms plan to report updated spending and productivity figures in the coming quarter.

Some investors and industry participants expect similar budget pressure could affect crypto and AI infrastructure projects that adopt comparable agentic workflows in the months ahead.