Enterprise

The token bill comes due: Microsoft's June 30 Claude Code cutoff caps an enterprise AI spending reckoning

Uber blew its 2026 AI coding budget by April, Meta pulled its tokenmaxxing leaderboard, and a $98M Stanford spinout just exited stealth promising to cut consumption 100x. The all-you-can-eat era is closing.

By Asher Glick · Enterprise software desk · 2026-06-26 · 3 min read

Photo: Unsplash / Mika Baumeister — Spreadsheet of financial figures open on a laptop, representing enterprise budget tracking

Microsoft has given engineers in its Experiences and Devices division four days to stop using Claude Code, with most internal licenses set to be cancelled by June 30, the last day of the company’s fiscal year. The Verge reported the cutoff in May; the teams behind Windows, Microsoft 365, Outlook, Teams, and Surface are being routed to GitHub Copilot CLI instead. The timing isn’t subtle. Fiscal year-end is when finance organizations stop tolerating line items that don’t pencil out.

The Microsoft decision lands in the middle of a broader correction. Uber burned through its entire 2026 AI coding budget in roughly four months, exhausting it by April. President and COO Andrew Macdonald, asked on the Rapid Response podcast whether soaring Claude Code usage was translating into shipped consumer features, said the link “is not there yet.” Uber had been running an internal leaderboard ranking teams by AI tool consumption. Meta quietly took down the informal tokenmaxxing leaderboard its own employees built. The Financial Times reported that some Amazon employees had been spinning up agents to perform meaningless tasks, padding usage stats ahead of manager review.

J.R. Storment, executive director of the FinOps Foundation, told TechCrunch that by April and May he was hearing from companies running 3x over their full-year 2026 token budgets. A Priceline employee said a Cursor renewal came back 4–5x more expensive. Per-engineer API costs of $500 to $2,000 a month are now common at comparable deployments. Ramp has moved into AI spend management; Datadog and New Relic have tacked on token-level observability and GPU monitoring. The vendor stack is reorganizing around the bill, not the capability.

On June 23, Engram, a 13-person Stanford spinout backed by General Catalyst, Kleiner Perkins, and Sequoia, exited stealth with $98 million and a $600 million valuation, pitching pre-trained KV cache memory it claims can cut token consumption 100x.

A year ago the leaderboards were the product. Now the guardrails are.

Sources