Falling AI Token Prices Are Fueling Demand, Not Killing It

Token prices keep falling, yet rental prices for Nvidia's H100 GPUs are holding steady or climbing. That is the core finding from new data published by venture firm Andreessen Horowitz (a16z), and it turns one of the AI market's biggest fears on its head.

The data, reported by The Decoder, points to a textbook Jevons paradox in the AI economy: as the unit cost of AI inference drops, consumption rises fast enough to more than offset the savings. Cheaper AI, in other words, is driving more demand, not less.

Why does this matter for Nvidia?

The pattern is Nvidia CEO Jensen Huang's best-case scenario. His argument, echoed across the AI supply chain, has long been that efficiency gains in AI models expand the market rather than shrink the hardware business behind it. The a16z numbers give that argument empirical backing.

If token prices fall but total token consumption grows even faster, compute demand keeps rising. That sustained demand is what keeps H100 rental prices from collapsing — and what keeps chipmakers, data center operators, and cloud providers investing.

What happens if demand flattens?

The a16z data carries a warning alongside the good news. The current equilibrium depends on demand continuing to grow faster than costs decline.

If demand flattens while token prices keep falling, the entire chain — from chipmakers to cloud providers — takes a hit. GPU rental prices would come under pressure, and the heavy capital spending built on assumptions of ever-growing inference demand would become harder to justify.

What is the Jevons paradox?

The Jevons paradox describes a situation where efficiency improvements reduce the cost of using a resource, but total consumption of that resource rises as a result. Cheaper access expands the range of economically viable uses.

In the AI market, the mechanism works like this:

• Model improvements and competition drive down the price per token.
• Lower prices make AI economical for new applications and higher-volume use.
• Total token consumption grows faster than the per-token price falls.
• Aggregate compute demand rises, supporting GPU prices.

The a16z data suggests the AI market is currently on the right side of this trade-off: falling token prices, rising total spend, and H100 rental prices that refuse to drop.

The stakes

The finding matters well beyond Nvidia's revenue line. The AI industry's capital cycle — data center construction, GPU procurement, cloud capacity commitments — rests on the assumption that cheaper inference expands the market. The a16z data supports that assumption for now, but it also identifies the precise condition under which the model breaks: demand growth flattening while costs keep falling.

The next signal to watch is whether H100 rental prices hold as newer GPUs enter the market and token prices continue to slide. If they do, Huang's best-case scenario holds; if they slip, the Jevons argument will face its first real stress test.