AI is getting cheaper, but computing power isn't following

The price of using AI is falling. But the price of renting the computer chips AI runs on is holding steady or rising. That's according to data from Ornn, Silicon Data and Bloomberg as of August 2026, as reported by The Decoder.

It may sound like a contradiction, but many see an explanation: When AI gets cheaper, we use more of it.

Tokens and GPUs in brief

A token is the small chunk of text an AI model reads or writes, and it is the unit AI is typically billed in. A GPU is the type of chip AI models run on. They are often rented out by the hour.

According to a Citadel Securities memo, usage-weighted token prices, as measured by Silicon Data, have fallen about 40 percent since the end of June 2026. Meanwhile, Ornn's measurements of GPU rental prices have risen again. The H100 chip's spot price (the daily price) fell 41 percent from its peak in May to its low in June, but has since climbed 20 percent. It is still below the May peak. The newer B200 costs about $6.17 an hour, which is above the May peak.

Jevons paradox

According to The Decoder, the investment firm a16z calls the trend a textbook case of Jevons paradox. The idea is that when something gets cheaper, consumption grows so much that total spending goes up. Cheaper tokens open the door to AI agents (systems that carry out tasks on their own), automation and new uses.

Citadel Securities believes that dynamic is becoming dominant. The bank had previously viewed it as secondary. According to the memo, AI spending per employee rose about 49 percent month over month in July among the top 1 percent of companies in Ramp's transaction data. Among the top 10 percent, the increase was 25 percent, and for the median it was 9 percent.

What if growth stops?

The signs are preliminary, and the period is short. The Decoder points out that it is unclear how much of the demand comes from people and how much the systems generate themselves, because agents use very large numbers of tokens. Demand may also be artificially inflated.

The model assumes that consumption grows fast enough to offset falling prices. As long as that happens, hardware remains scarce and expensive. If demand levels off, it could hit the entire chain, from chip and memory suppliers to energy companies and cloud providers, and markets could spiral.

That sensitivity was on display when US stocks fell on reports that OpenAI's annualized revenue could be lower than previously reported.