Falling AI token prices drive up H100 GPU demand

Data through August 2026 from Ornn, Silicon Data, and Bloomberg reveals a textbook Jevons paradox unfolding across the artificial intelligence sector, according to venture firm a16z. While the cost of processing individual AI tokens continues to drop sharply, rental prices for Nvidia H100 GPUs are holding steady or even climbing. The decline in unit costs has not dampened revenue for infrastructure providers because cheaper tokens are expanding the broader market at a faster rate than per-unit prices fall.

This surge in compute demand is largely propelled by the rapid rise of agentic AI, automated workflows, and novel software applications. Autonomous agents consume tokens at an extraordinary pace compared to traditional human interaction, creating massive backend compute requirements. Consequently, even modest growth in human end-users can trigger outsized demand for hardware resources, keeping data center capacity scarce and high-performance hardware consistently expensive.

The entire AI infrastructure pipeline relies heavily on the premise that total token volume will expand quickly enough to compensate for plummeting per-unit pricing. If this demand growth slows or flattens, the financial impact will ripple rapidly across the technology ecosystem, affecting chip manufacturers, memory suppliers, energy providers, and cloud computing firms.

The market remains remarkably sensitive to any signals that software adoption might fail to keep pace with hardware investments. This underlying fragility was recently underscored when US stock markets tumbled following reports that OpenAI generated lower annualized revenue than previously projected. For now, hardware providers remain in an enviable position, but their continued success depends entirely on software systems maintaining an insatiable appetite for compute.