Cheap AI tokens keep GPU rentals high in a textbook Jevons paradox

AI token prices have fallen for two straight years, but Nvidia H100 GPU rental rates have not followed. Data compiled by Ornn, Silicon Data, and Bloomberg through August 2026 shows the pattern: cheaper tokens are driving total demand higher, not lower. Venture firm a16z calls it a textbook case of Jevons' paradox - when a resource becomes cheaper, consumption rises so much that overall demand increases rather than decreases.

For Nvidia and CEO Jensen Huang, this is the scenario they have bet on since the ChatGPT boom began. Cheap tokens make AI agents and automated workflows economically viable in ways they were not a year ago, pulling in more usage and keeping GPU capacity tight even as per-token costs collapse.

The measurement problem no one has solved

The data cannot cleanly separate human-driven demand from machines querying machines. Agentic systems chew through tokens at rates that dwarf anything a person typing into a chatbot produces. That ambiguity matters because it determines whether the hardware shortage propping up chipmakers, memory suppliers, and cloud providers reflects real growth or something more fragile.

If most of the new volume is agents burning tokens to accomplish tasks rather than people doing new things, the Jevons framing starts to look less like organic market growth and more like a measurement artifact. Founders building on top of these token economics should ask whose usage is actually paying the bills.

Why the gap persists

H100 rental rates have held steady or climbed because cheaper tokens unlocked new AI applications, automation systems, and agentic workflows that require massive computational resources. The increased volume of AI workloads driven by lower per-token costs means demand for GPUs continues to grow, preventing rental prices from declining alongside token costs.

For Nvidia, this is the ideal outcome. Cheaper tokens did not kill the GPU business - they expanded it. Steady H100 rental prices while token costs fall is the clearest signal yet that demand, not just capacity, is driving the market. Researchers tracking compute trends should want the human-versus-agent split before citing this as proof that cheaper AI reliably creates proportionally larger markets. Right now, that is assumed, not shown.

Why this matters for executives and strategy leaders

The token-to-GPU pricing gap creates a planning problem for any organization budgeting AI infrastructure. If agentic systems are the primary driver of volume growth, cost models built on human usage patterns will break. Teams building automated workflows should model for scenarios where token consumption grows faster than user adoption, and procurement should question whether fixed GPU contracts reflect real demand or a temporary measurement blind spot. Leaders pursuing AI Technology Leadership Courses can build the technical fluency needed to pressure-test these assumptions before committing capital.