The "AI is almost free" phase is over

Last month I wrote that the "AI is almost free" phase was ending. This article is the receipt.

Microsoft reportedly canceling most Claude Code licenses. Uber burning through its full 2026 AI coding tools budget in four months. Nvidia's own deep learning VP saying compute now costs more than the people on his team.

And the math is even harder outside the US.

Tokens cost the same everywhere. Engineers do not. In Bangalore, Manila, Jakarta, or Kuala Lumpur, an agentic workflow can quietly cost more than the engineer it was meant to help.

When the subsidy layer goes, "use more AI" stops being a strategy. Token leaderboards stop making sense.

The next phase is deliberate AI. Routing between models. Smaller models for simple tasks. Private specialised models for repeatable workflows. Frontier models only where they earn their cost.

Companies will need people who understand where AI creates value and where it only burns tokens, and metrics to back it up. Token usage per task. Cost per feature. Rework rate. ROI by workflow.

AI is not going away. It is becoming normal engineering infrastructure. But infrastructure comes with a bill.