AI Token Deflation Masks a Two-Tier Market with Widening Margins

Key points

• Silicon Data’s Token Expenditure Index fell 52.4% from a late-May high of $2.0651 per million tokens, driven by efficiency gains and price wars.
• Epoch AI reports inference costs for equivalent performance have dropped 47% per quarter since 2023, outpacing historical tech advancements like DNA sequencing and battery tech.
• Seaport Research Partners notes a two-tier market: US frontier labs maintain ~70% gross margins, while Chinese open-weight operators operate at ~20% margins.
• The median token price has stabilized near $1 per million since mid-2024, with a floor price of $0.11 reflecting electricity costs, despite falling averages.
• Distillation techniques allow smaller models to mimic frontier performance, accelerating cost declines and compressing the window for premium pricing by leading labs.

Background

This trend follows earlier reports in September 2026 where token prices hit fresh lows due to OpenAI’s price cuts and increased competition from open-source models. The current analysis provides a deeper structural view of these price drops, moving beyond simple average metrics to examine margin differentials and production costs.

How outlets are covering it

The Financial Times emphasizes that the blended price index is misleading because it conflates technical efficiency, mix shifts, and competitive pricing. In contrast, Chosun Ilbo highlights the sheer speed of cost reduction, comparing it to historical industrial breakthroughs. Citi’s perspective, noted in secondary sources, suggests frontier models are widening their lead over open-weight rivals, contradicting the notion of pure commoditization. While all sources agree on falling costs, they diverge on whether this signals a margin war or a sustainable two-tier market structure.

Why it matters

Investors and enterprises must understand that falling token prices do not necessarily mean collapsing revenues for AI providers. The stability of gross margins for frontier labs suggests that despite price deflation, the business model remains robust. However, the rapid decline in inference costs could accelerate the adoption of AI agents and reshape competitive dynamics between US and Chinese AI ecosystems.

What to watch

Watch for potential regulatory or protective measures by national governments to safeguard domestic AI leaders, as suggested by the Financial Times. Additionally, monitor the impact of distillation techniques on the performance gap between frontier and open-weight models, which could further compress margins for mid-tier providers.