AI: Token Prices Drop, GPU Farms, Demand-Dependent Balance

Token prices are declining, but H100 GPU rentals remain strong. Data extending to August 2026 describes a dynamic akin to a Jevons Paradox, where usage rises faster than costs decrease. This balance benefits those selling compute as long as adoption accelerates, but it exposes the ecosystem if demand wanes. The stock market nervousness related to doubts about OpenAI's revenues has provided a glimpse into this.

Markets Already Anxious About Demand Slowdown Risks

U.S. stocks fell after reports suggested that OpenAI's annualized revenues could be lower than previous indications. This type of signal fuels fears of a downturn, with some estimating that markets could collapse if demand stabilizes. In this scenario, the chain linking chip manufacturers, memory suppliers, energy providers, and cloud companies could suffer. Conversely, as long as AI usage grows enough to offset the decline in token prices, hardware remains scarce and expensive, which is the central hypothesis at the moment.

Token Prices Drop While H100 GPU Rentals Hold Steady

Token prices continue to decrease, while rental rates for H100 GPUs remain stable or even increase. Datasets from Ornn, Silicon Data, and Bloomberg, available until August 2026, document this divergence, including a comparison between the token price index and the rental cost of H100s. a16z describes this regime as a "Jevons Paradox" applied to AI, and it is reported that cheaper AI tokens stimulate demand.

Token-Hungry Agents and Uncertainties About Demand Sources

The share of demand attributable to human users versus systems is not clearly established, as agentic AI consumes very high volumes of tokens. According to some, a decrease in token prices would encourage the development of agents, automation, and new use cases, leading to a growth in usage volume that outpaces the reduction in cost per token. It is possible that demand for compute capacity is artificially amplified, and even a limited increase in human activity could be enough to create significant hardware needs. This context is described as the most favorable scenario for Jensen Huang.