Enterprise AI Coding Costs: The Microsoft and Uber Wake-Up

When the Tool Succeeds Too Well

In April 2026, Uber's Chief Technology Officer for Mobility and Delivery, Praveen Neppalli Naga, confirmed to The Information that the company had exhausted its entire 2026 AI coding tools budget — four months into the fiscal year. The problem was not that the tools failed. It was that they worked: Claude Code usage within Naga's approximately 5,000-engineer organisation climbed from 32 per cent of engineers to 84 per cent in a single month by March 2026. Individual engineers were spending between 500 and 2,000 US dollars per month on tokens. The budget gave out before the financial year was halfway through.

Uber's situation is the clearest public example of a structural problem that has emerged at the top of enterprise engineering organisations in the first half of 2026: AI coding tools are good enough that engineers use them constantly, and constant use at frontier model token rates breaks the economics of fixed engineering budgets.

Microsoft Follows: Experiences + Devices Out of Claude Code by 30 June

Within weeks of Uber's disclosure, a second major story landed. Microsoft's Experiences + Devices division — the organisation responsible for Windows, Microsoft 365, Outlook, Teams, and Surface — is cancelling most Claude Code licenses across the division by 30 June 2026. Engineers are being directed to switch to GitHub Copilot CLI, Microsoft's own command-line AI coding tool.

The timing is not coincidental. Microsoft's financial year ends on 30 June, which means the cancellation closes an AI expenditure line before it crosses into FY27. In December 2025, Microsoft had extended Claude Code access to thousands of engineers, product managers, and designers. By spring 2026, the tool had spread well beyond the engineering teams it was initially scoped for — non-technical roles with no awareness of token pricing had adopted it freely, accelerating spend beyond what the budget could absorb.

The Structural Problem With Metered Billing

Both the Uber and Microsoft situations share the same architecture of failure. AI coding tools under flat-fee subscription models carried predictable monthly costs regardless of usage intensity. The shift to token-based metered billing — which GitHub Copilot itself enacted on 1 June 2026 — turned that predictable overhead into a variable cost that scales linearly with how well and how often engineers use the tool.

An engineer who uses an AI coding assistant occasionally for inline completions costs tens of dollars a month. An engineer who runs agentic coding sessions for hours daily, routes large-context tasks to frontier models, and has the tool handle test writing, review cycles, and documentation, costs hundreds to thousands of dollars a month at current token prices. Multiply the latter pattern across thousands of engineers, and the monthly bill exceeds any budget that was written when the tool was treated as a convenience utility rather than core infrastructure.

The Metrics Question

Black Duck's State of AI-Powered Software Development report, published in 2026, found that AI coding adoption across enterprise engineering teams had reached 97 per cent. Lines of code added per developer per week roughly doubled year-over-year. Pull requests with more than 1,000 lines of changes became notably more common from January 2026 onward.

Those are velocity metrics. What they do not measure is whether the acceleration in code volume corresponds to lower defect rates, fewer customer-reported bugs, or faster time to measurable business outcomes. Uber's COO's public comments after the budget incident included precisely this concern: the company could observe usage climbing, but not whether the usage translated into business value at a rate that justified the cost.

The Stack Overflow 2025 developer survey found that more developers actively distrust AI coding tools — 46 per cent — than trust them — 33 per cent. That gap is not narrowing on its own.

What This Means for Indian Engineering Teams

The Uber and Microsoft situations are a calibration signal for every engineering organisation scaling AI coding tool adoption in 2026. Three practical adjustments apply.

First, treating AI coding spend as a fixed overhead is now structurally incorrect. Token-based metered billing means costs scale with usage intensity, not headcount. Teams that track spend at the project or team level — not only at the company level — will catch budget overruns before they reach the end of a fiscal quarter.

Second, tool adoption governance matters more than it did under flat-fee models. Extending AI coding tool access to non-technical roles without token-consumption training creates the exact spending pattern that accelerated Microsoft's budget problem — high-intensity usage by people with no mechanism to evaluate whether each session's cost is justified.

Third, the open-source, bring-your-own-key architecture of tools like OpenCode becomes more economically attractive precisely in this environment. With bring-your-own-key tools, cost per task is bounded by the API rate you choose, not by the premium a managed service charges. Teams that route simpler work to cheaper or local models — and reserve frontier API spend for genuinely hard tasks — have a structural cost advantage over teams that default to frontier models for every request.

The Bottom Line

Uber exhausted its 2026 AI coding tools budget in four months as per-engineer token spend reached 500 to 2,000 US dollars monthly. Microsoft cut Claude Code for its Experiences + Devices division by 30 June 2026. Both cases share the same structure: AI coding tools scaled beyond budget models that assumed flat-fee software. For Indian engineering teams, the message is direct — AI coding spend is now a variable infrastructure cost requiring the same governance discipline as cloud spend, not the same treatment as a software subscription.