Will falling cost per token increase aggregate accelerator spending or reduce it?

Evidence(2)

1. 

   July 2026: Token prices fell while enterprise AI spending and GPU rental prices remained firm

   [2]

   Jul 2026

   Citadel’s data as of July 2026 showed usage-weighted token prices declining 40% since the end of June, while AI spend per employee among the top 1% of enterprises increased by approximately 50% month over month in July. GPU rental markets remained firm relative to their June lows, with rental prices reportedly rebounding while token prices continued to decline. These observations are consistent with a possible usage rebound, but they do not establish an industry-wide demand elasticity or prove that additional workload growth outweighed efficiency gains.

   1

   Citadel’s July-2026 observations provide the report’s clearest near-term demand-response signal: usage-weighted token prices fell while enterprise AI spending rose and GPU rental pricing remained firm. The report treats this as tentative support for a rebound in underlying compute demand, not proof that incremental demand exceeded efficiency gains. The evidence is concentrated among leading enterprises, and the rental market is fragmented and opaque.

   See how this bears on the question
2. 

   NVIDIA reported Q2 FY2027 results

   [4]

   26 Aug 2026

   NVIDIA reported results on August 26, 2026, for the quarter ended July 26, 2026. Revenue was US$96.2 billion, up 106% year over year, including US$89.0 billion of Data Center revenue, up 117% year over year. GAAP and non-GAAP gross margins were both 75.0%. The report uses these results as evidence of substantial upstream demand, while noting that company revenue is not a measure of aggregate accelerator spending or a clean measure of inference-only demand.

   See how this bears on the question