Falling token prices are not reducing GPU demand, a16z data shows

By Splitfeed EditorialLast updated Oct 11, 2026, 10:00 UTC

On October 11, 2026, venture capital firm Andreessen Horowitz published data tracking two pricing trends in the AI market simultaneously: the cost of AI tokens has continued to fall, while rental prices for Nvidia H100 GPUs have held steady or moved higher. The firm's analysis frames this as a textbook instance of the Jevons paradox, the economic pattern in which efficiency gains lower the cost of a resource enough to stimulate demand beyond what those gains save, resulting in greater total consumption rather than less.

The significance here goes beyond a single data point about chip prices. If the a16z figures hold up, they suggest that the AI industry's cost curve is not translating into reduced infrastructure spending, which undermines a common assumption that falling inference prices would eventually cool the GPU buildout. For Nvidia and the cloud providers renting H100 capacity, this is the most favorable possible interpretation of current market dynamics. The genuine dispute is whether the demand growth sustaining those GPU prices is durable or whether it reflects a wave of experimentation and deployment that could plateau. The data also implicitly raises questions about who captures the value when token prices fall: end users and application builders benefit from cheaper inference, but if GPU prices stay firm, the gains do not necessarily flow back to reduce infrastructure costs for the broader industry.

With no published reactions yet from the Pro-AI, Anti-AI, or Middle Ground camps on this specific story, the likely lines of argument can be sketched by disposition. The Pro-AI camp would typically treat the Jevons dynamic as confirmation that AI adoption is genuinely accelerating and that demand is organic rather than artificially inflated. The Anti-AI camp would be expected to focus on the downside scenario a16z itself flags, that a flattening of demand growth would send a damaging chain reaction through chip makers and cloud providers, raising questions about overinvestment. A Middle Ground position would likely acknowledge both sides, accepting the paradox as real while treating the stability of demand growth as the key unresolved variable.

The argument will sharpen around any forthcoming earnings reports from Nvidia and major cloud providers, which would show whether H100 utilization rates and rental revenues are tracking the pattern a16z describes or beginning to soften. Analyst guidance from those companies in late 2026 would offer the clearest near-term test of whether the Jevons dynamic is holding.