For months the market measured success by usage. But usage and money stopped moving together, and that separation changes the whole story.
Published on July 4, 2026 · by Alex · Artificial intelligence · Margins · Data centres
A signal appeared that very few are watching and that could change the sector's story: the indicator tracking spending on AI tokens — the money companies pay to use these models — fell around 20% from its late-May peak.
It is worth reading carefully, because it does not mean what it looks like. It is not that artificial intelligence is used less. It is that customers found a way to use the same thing for less money.
Far cheaper models appeared, especially from China, capable of handling much of the work for a fraction of the cost. And when a company gets similar results paying much less, the first casualty is always the same: pricing power.
That is what separates this signal from ordinary noise. A sector can survive growing more slowly. What is much harder to survive is unit prices collapsing while investment keeps climbing.
If revenue grows more slowly, the question gets uncomfortable: how do you justify the hundreds of billions going into data centres, chips and new infrastructure.
Some analysts already compare the gap between investment and sales to the overspending in telecoms before the year 2000. The parallel has a limit — back then fibre was laid that took a decade to use, while today capacity is occupied almost as soon as it is switched on — but the underlying arithmetic is the same: spending running ahead of revenue.
This story has an honest counterweight. Demand for memory and chips is still enormous, the most advanced processors remain effectively sold out, and slack is not expected until 2028. Many companies keep expanding their projects.
In other words: the problem is not a lack of customers. It is that those customers now want more efficient, cheaper models. It is a margin problem, not a demand problem — which is worse than it sounds, because a demand problem can be seen coming and a margin problem shows up directly in the results.
On top of this come new regulations in the United States and Europe, which make the most powerful models more expensive and slower to deploy. The practical effect is to push many companies toward lower-cost alternatives — exactly the same direction as the price pressure.
The real competition is no longer who builds the best artificial intelligence. It is who manages to make money from it. If customers stop paying premium prices, the sector's next great challenge will not be technological: it will be financial.
And that is the figure likely to define the next winners and losers, far more than any new model announcement.
No. What fell roughly 20% from its late-May peak is token spending — the money companies pay to use the models. Usage is not down: customers found ways to do the same thing for less, largely with far cheaper Chinese models.
Because a demand problem can be seen coming, while a margin problem shows up straight in the results. Memory and chip demand is still enormous and the most advanced units are still sold out; what is eroding is the pricing power of whoever sells the models.
The underlying arithmetic is the same: spending running ahead of revenue. The parallel has a real limit — back then fiber was laid that took a decade to use, while today capacity is filled as soon as it is switched on — but the gap between investment and sales rings the same bell.
This article is the written version of the Saturday analysis.
Watch the video on TikTok · Follow the channel
