The LLM Token Index fell from approximately $2.05 per million tokens on May 28, 2026, to $0.98 as shown in the October chart, a drop of about 52%. However, during the same period, the rental price indices (H100, B200, and B300, with May 31 as 100) increased by 3.3%, 10.2%, and 38.6%, respectively. The decrease in token unit cost coincided with the increase in GPU rental prices. This may reflect improved model inference efficiency, increased competition between models, and an increased supply of low-priced tokens, driving down the unit cost of tokens; simultaneously, the continued growth in AI computing power demand supported GPU rental prices. This divergence between the two sets of indicators means that a decrease in the unit cost of model services does not necessarily imply a simultaneous easing of the supply and demand for underlying computing power.
