In this latest AI selloff in US stocks, I think the market is recalculating a rather ugly set of numbers.
Data from Citadel Securities at the end of October showed two figures that really stand out when put side by side.
Over the past six months, the price the market actually paid for tokens fell by about 50%, while the average price of frontier models dropped 62%. At the same time, rental prices for B200 GPUs rose 27.5%.
The products being sold are getting cheaper, while the things needed to produce them are getting more expensive.
According to Citadel’s calculations, all else being equal, model companies would have to increase their token output per hour by 2.55 times just to maintain their previous gross margins.
I did the math myself:
1.275 ÷ 0.5 ≈ 2.55.
It lines up exactly with those two figures above.
It’s like the situation manufacturing bosses fear most: factories and equipment keep getting more expensive, while the prices of finished goods keep falling. And these US companies are still borrowing money to buy GPUs, build data centers, and sign long-term power contracts.
And there’s more pressure on prices from below.
On OpenRouter, Chinese models’ share of tokens went from 6%–13% in February this year to 57%–67% in September.
They don’t have to win at everything.
For tasks like customer service, translation, and content generation, if the results are good enough and the price is 60% to 90% lower, companies have an incentive to switch.
Being number one on the leaderboard doesn’t solve this problem. Token prices have already been driven down.
The market used to assume that the more tokens people used, the more money model companies made, which is why it was willing to assign such high valuations to GPUs, cloud services, and data centers.
Now usage is still rising, but the money users pay isn’t rising along with it.
Demand for computing power may continue to grow, and Nvidia’s cards may keep selling. But the books are going to get harder and harder to balance for foundation model companies, cloud providers facing shrinking margins, and all those companies borrowing money to build data centers.
When I look at AI companies going forward, I won’t just look at how many tokens they process. I’ll want to know how many dollars those tokens ultimately bring in.
Usage is growing fast, but revenue isn’t keeping up.
Someone will eventually have to foot the bill for hundreds of billions of dollars in data centers.
Data from Citadel Securities at the end of October showed two figures that really stand out when put side by side.
Over the past six months, the price the market actually paid for tokens fell by about 50%, while the average price of frontier models dropped 62%. At the same time, rental prices for B200 GPUs rose 27.5%.
The products being sold are getting cheaper, while the things needed to produce them are getting more expensive.
According to Citadel’s calculations, all else being equal, model companies would have to increase their token output per hour by 2.55 times just to maintain their previous gross margins.
I did the math myself:
1.275 ÷ 0.5 ≈ 2.55.
It lines up exactly with those two figures above.
It’s like the situation manufacturing bosses fear most: factories and equipment keep getting more expensive, while the prices of finished goods keep falling. And these US companies are still borrowing money to buy GPUs, build data centers, and sign long-term power contracts.
And there’s more pressure on prices from below.
On OpenRouter, Chinese models’ share of tokens went from 6%–13% in February this year to 57%–67% in September.
They don’t have to win at everything.
For tasks like customer service, translation, and content generation, if the results are good enough and the price is 60% to 90% lower, companies have an incentive to switch.
Being number one on the leaderboard doesn’t solve this problem. Token prices have already been driven down.
The market used to assume that the more tokens people used, the more money model companies made, which is why it was willing to assign such high valuations to GPUs, cloud services, and data centers.
Now usage is still rising, but the money users pay isn’t rising along with it.
Demand for computing power may continue to grow, and Nvidia’s cards may keep selling. But the books are going to get harder and harder to balance for foundation model companies, cloud providers facing shrinking margins, and all those companies borrowing money to build data centers.
When I look at AI companies going forward, I won’t just look at how many tokens they process. I’ll want to know how many dollars those tokens ultimately bring in.
Usage is growing fast, but revenue isn’t keeping up.
Someone will eventually have to foot the bill for hundreds of billions of dollars in data centers.
