AI Keeps Getting Cheaper, but GPU Rental Prices Won't Budge: Inside a16z's Analysis
a16z argues that as AI gets cheaper to use, compute demand is going up rather than down. With rental prices holding even for older GPUs, the open question is whether that extra demand sticks.
In “State of Markets II,” a market analysis report published on September 30, the venture firm a16z argued that falling AI usage costs and rising compute demand are happening at the same time. The price of tokens—the small chunks of text that serve as the billing unit for what goes into an AI model and what comes back out—keeps dropping, yet rental prices for the GPUs that run those models are holding steady or climbing.
Why GPU rental prices hold up even as AI gets cheaper
The report’s central point is that total compute demand can grow even when the cost of a given unit of AI work falls. a16z’s reading is that cheaper AI drives more usage, and that demand is growing faster than costs are coming down. What users pay to use AI and what service operators pay to rent compute are moving in opposite directions.
In an October 11 report, The Decoder framed this as the “Jevons paradox”: when efficiency gains lower the cost of using a resource, consumption rises and overall demand ends up larger. Citing a16z’s data, the outlet said rental prices for H100 GPUs are holding steady or rising. The implication is that cheaper AI translates into stronger demand for hardware.
Rental prices for older GPUs were another point the report highlighted. There had been concern that rental rates for previous-generation hardware would slide over time. a16z said that at the time of observation, A100 rental prices were at or above where they stood at the start of the year. The report cited both rising rental prices for newer-generation GPUs and steady prices for older hardware as evidence of compute demand.
Will the extra demand hold?
a16z sees paid AI usage as still being in its early innings. According to the report, as of April 2026 roughly 2% of US households pay for an AI service. a16z treats that figure as evidence that AI adoption has plenty of room left to spread.
Among companies, there was a gap between talking about AI’s impact and reporting it in numbers. Per a16z, about 30% of S&P 500 companies mentioned a quantified impact from AI. Only about 2%, by contrast, reported metrics they actually track. The report concludes that enterprise AI adoption is also at an early stage.
The Decoder characterized this demand expansion as a favorable scenario for Nvidia’s Jensen Huang. At the same time, it noted that if demand stalls, everyone from chipmakers to cloud providers could take a hit. Ultimately, the focus narrows to two questions: whether falling AI usage costs keep translating into rising usage, and whether GPU rental prices continue to bear that demand out.