Recent moves from major AI players signal the end of the era of cheap, flat‑rate AI. Microsoft revoked internal Claude Code licenses, Uber exhausted its 2026 AI budget in months, and GitHub scrapped flat‑rate plans. Companies assumed inference costs would keep falling, but usage spikes and rising hardware prices have shattered that assumption, forcing a pricing rethink. The shift reveals hidden operating costs.

Supply constraints now dominate cost curves. Memory prices have quadrupled and GPU prices have risen about two‑fold, and cooling constraints, for data centers, with high‑bandwidth memory driving a 435 % increase in bill‑of‑materials for Nvidia’s new VR200 chips. Anthropic’s CFO testified the firm spent $10 billion on compute while generating $5 billion in revenue, illustrating that labs are operating at a loss on inference.

Product teams must now price by usage rather than by seat. Per‑action billing ties revenue to each API call, while prepaid credit buckets smooth cash flow across multiple models. A hybrid seat‑plus‑credits scheme has emerged as the default for enterprise contracts for SaaS. Aligning pricing with real hardware costs lets companies keep building AI features without eroding margins.