‘The AI bubble might burst quicker than I thought.’

by

Tech giants spin efficiency narrative and hide the brutal inference expense reality.

Brian Roemmele
@BrianRoemmele

MICROSOFT CANCELED CLAUDE CODE!

IT COST TOO MUCH.

Major tech companies are confronting the steep reality of AI inference costs as the era of heavy subsidies appears to be ending.

Microsoft is canceling most internal licenses for Anthropic’s popular Claude Code tool by June 30, 2026 less than six months after rolling it out broadly to engineers primarily due to escalating token-based expenses, while shifting teams toward its own GitHub Copilot CLI.

This mirrors broader pressures: Uber’s CTO revealed the company had already exhausted its entire 2026 AI budget in just four months thanks to heavy Claude usage among thousands of engineers (with individual monthly costs often hitting $500–$2,000), and GitHub is transitioning Copilot to usage-based billing with higher per-token rates starting June 1st.

The reality is good enough AI will expand and constantly get better removing the oxygen of the most expensive.