The price of AI tokens is collapsing: towards a commoditization of the market?

The software was to devour the world and reap XXL margins in the process. Fifteen years later, it is the software itself that is starting to be devoured, by its own prices. The effective price paid by US companies for a million AI tokens has collapsed 41% since its peak in March, from $1.15 to 68 cents.

The token, a new interchangeable raw material

According to data from Ramp, the corporate spending platform, published on September 9, the detail by actor obtained by Fortune gives the measure of the movement. Since August 1 alone, the effective price of OpenAI has fallen 38%, to 48 cents. Anthropic is falling a little slower, 22% down, to 90 cents.

Another signal which is more worrying than the drop in prices itself: the share of use captured by the so-called “frontier” models, the most powerful and most expensive on the market, slips from 53% at the beginning of August to 45% in September. Businesses migrate to cheaper models as soon as the job allows, and no longer have much reason to pay the premium rate by default.

In comments reported by Fortune, Ara Kharazian, chief economist at Ramp, sums up the dynamic in a catchy phrase: tokens are starting to behave like salt or wheat, an interchangeable raw material, rather than a differentiating product that warrants a brand premium. An observation which is a stain for companies whose valuation is largely based on the idea that their models will remain irreplaceable.

One more crack in the AI bubble thesis

This is not an isolated accident, it is the direct consequence of increasingly fierce competition between laboratories, confirmed at the beginning of September by CNBC, to which is added the rise in power of Chinese models with open weights, significantly less expensive to run.

Le Journal du Coin was already dissecting, at the end of August, the signals of a slow deflation of the AI ​​bubble rather than a brutal crash like 2000. This fall in the price of tokens is part of the same mechanism: no spectacular collapse, but a slow erosion of margins which eats away, month after month, the central argument sold to investors.

The paradox is that this drop occurs even as Anthropic prepares an IPO reported to be worth more than $2,000 billion, and OpenAI continues to raise tens of billions to finance its own infrastructure. More affordable tokens should, in theory, boost adoption and therefore volumes. But higher volume at lower margins does not necessarily yield more, and there is no guarantee that the decline in inference costs will follow the same pace as the price reduction imposed by the competition.