Fortune's recent piece argues that falling AI inference prices will not deliver corresponding savings on enterprise AI bills, framing a divergence between unit economics and total contract value that now defines corporate AI procurement.

The thesis hinges on a structural gap: per-token model costs continue to drop, but enterprise AI spending does not follow the same curve. Three forces drive the divergence.

Why do per-token prices keep falling?

Major model providers compete aggressively on flagship and distilled tiers. New releases generally price below predecessors, and dedicated mini models address cost-sensitive workloads. Frontier capability still commands premium rates, yet the directional pressure across the market pushes unit costs downward. Inference hardware improvements reinforce the trend.

Why do enterprise bills keep climbing?

Enterprise AI contracts bear little resemblance to public API price lists. Buyers typically commit to:

• Reserved throughput or seat-based agreements with monthly floors
• Premium reasoning-tier access priced per seat rather than per token
• Higher rates for extended context windows used in document analysis
• Storage, retrieval pipelines, and vector database line items
• Deployment, integration, and managed-service fees

Volume rises faster than per-call price falls. A legal team moving from a 50-seat pilot to a 1,000-seat rollout multiplies usage even as per-call costs halve. Many enterprises have moved past pilots into scaled production within the past 18 months, broadening the user base that any unit-cost decline must serve.

What's driving the agentic multiplier?

Multi-step agent workflows amplify the gap. A single API call once resolved a user query. An agentic system chains reasoning steps, tool calls, and verification passes — each consuming tokens at the model's premium reasoning rates. The piece flags this dynamic as central to the divergence: gains in autonomy convert directly into proportional cost expansion.

How do vendor economics shape customer outcomes?

Model providers carry heavy capital expenditure from the GPU buildout cycle and pursue gross-margin recovery through enterprise tiers. Multi-year agreements include escalator clauses tied to feature releases and minimum-commitment penalties that bind customers to spend floors regardless of throughput optimization. Buyers rarely achieve the headline API price in practice.

What are procurement teams doing?

CIOs and procurement leads have begun modeling AI spend differently:

• Renegotiating reserved-capacity terms as usage patterns mature
• Routing simple queries to smaller, cheaper models in multi-model architectures
• Implementing internal chargeback so business units own their AI consumption
• Insisting on usage dashboards as enterprise deal prerequisites

What remains unresolved?

The piece does not project when, or if, the gap closes. Efficiency breakthroughs, market saturation, or board-level ROI scrutiny could compress margins or constrain demand. No scenario guarantees an imminent resolution.

For now, the warning to AI buyers is concrete: cheaper inference does not equal cheaper AI bills. Enterprises building multi-year AI budgets should model the divergence between token-cost trends and total contract value rather than assume savings flow automatically.