Token prices fall, GPU rental stays high: Jevons in the AI market

Token prices have been falling for years. Yet renting an H100 today costs no less, and probably more. The Decoder reported on 11 October 2026 on a post by the venture firm a16z on X, and cloudmagazin the same day on a16z material. Both are secondhand. That matters for how to read it, because the core claim sits in a chart that the text does not spell out in numbers.

The argument concerns compute time on NVIDIA accelerators. Neither report gives concrete prices per token or per GPU hour. (Quelle: NVIDIA)

What a16z shows

In The Decoder's description, the chart sets the H100 rental price against a token price index, with data through August 2026 from Ornn, Silicon Data and Bloomberg. Falling token prices contrast with H100 rental prices that stay high or rise. The post calls this a classic Jevons paradox: cheaper tokens make agents, automation and new applications economical, and volume grows faster than unit costs fall.

According to The Decoder it is unclear how much demand comes from people and how much the systems generate themselves. Agents consume tokens very wastefully. (Quelle: The Decoder, 11 October 2026)

The report states the assumption behind it: usage has to grow fast enough to offset falling token prices. If demand flattens, the whole chain is hit, from chip makers and memory suppliers to utilities and cloud providers. On the market side, The Decoder points to a negative US market reaction to reports that OpenAI's projected annual revenue may be lower than assumed.

What cloudmagazin adds in numbers

The magazine summarizes a16z material: David George of a16z Growth in "State of Markets II" (30 September 2026), the Machine Age Fund hardware note (28 August) and Vanguard, which derived capex from Bloomberg data (as of 31 July). Combined capital spending of the five big cloud operators (Alphabet, Amazon, Meta, Microsoft, Oracle):

Euro values at the ECB reference rate of 30 September 2026 (1.1355). 2026 and 2027 are Bloomberg consensus estimates and no commitments by individual operators. (Quelle: cloudmagazin, 11 October 2026)

Other figures from the report: rack density rose 28-fold from the H100 rack to the Rubin rack, and systems in use draw 100 to 250 kW per rack. A100 rental prices sit at or above their level at the start of the year, though the text gives no hourly rates. George says the hyperscalers' free cash flow is becoming the semiconductor firms' free cash flow, funded by profits and increasingly by debt. According to the report, networking inside the racks is reaching the limits of copper cabling.

What this means for renting versus buying

The argument supports planning that does not write off older cards. If tokens get cheaper and demand for compute rises anyway, rental prices and resale values of older accelerators do not fall automatically. It is not established, because the numbers live in a chart we have not seen ourselves. The direction says something, the magnitude cannot be read off.

We run our own servers with RTX 5090 cards: one with four cards. The a16z finding still does not settle rent versus buy for us, because it concerns H100 data-center cards, not consumer GPUs. For your own calculation three figures count: how many hours a day the cards really compute, what power and cooling cost, and who looks after them. A rented H100 costs money every hour, load or not. A bought card costs only idle power without load, but the purchase and the risk of model changes are fixed.

The report also carries a caution: the paradox holds only while growth continues. Cloudmagazin calls the 2026 and 2027 figures estimates. Anyone committing long term, by lease or by purchase, should run both scenarios, including the one where demand flattens.