The AI Buildout Has Priced the Factory and Not the Market. 21bn Devices Have No Brain Yet.
Nvidia's data centre business grew 110% this quarter. If that sounds like a bubble, look at what the industry has actually installed: 21bn connected devices, and almost none of them have a brain yet.
What the quarter actually said
Nvidia sold $115.8bn of hardware in the three months to 25 October, up 103% from a year earlier. Data centre — the chips that go into AI machines — was $107.6bn of that, up 110%. Its gross margin, the share of revenue left after the cost of the goods, was 74.9%. Earnings per share were $2.68.
For the next quarter it guided $130.0bn, ahead of what analysts expected, with the gross margin falling to 71–72%. It also said, in a phrase almost nobody quoted, that the period contains **one extra week**.
That extra week is the whole story. The September quarter had 13 weeks; the January quarter has 14. Revenue per week therefore goes from $8.91bn to $9.29bn — a rise of 4.2%, not the 12.3% the headline comparison suggests. Gross profit per week is $6.67bn against a guided $6.59–6.69bn: lower at the bottom of the range, slightly higher at the top. Operating profit per week, after running costs, falls about 2% at the midpoint and is lower across the entire guided range.
None of that is a collapse. It is what a company looks like when it is installing a brand-new generation of equipment into factories and giving up some margin to do it. But a business whose profit per week is flat while its revenue grows 4% is no longer accelerating, whatever the headline says.
Why 2000 is the wrong comparison
In 2000 the buildout was fibre-optic cable and routers, sold to telecom companies, paid for with borrowed money. The end product — a web page — had no price of its own. When the telecoms stopped buying, there was no second customer, and the capacity sat idle for a decade.
Today the customer list is different in kind, and the counts make the argument. 21.1bn connected devices were in use at the last count, with 21.9bn forecast for 2026. 5.1 million industrial robots are working in factories, with just over 603,000 installed last year and 655,000 forecast this year. AI-capable personal computers should be 52% of shipments in 2026, roughly 143 million machines.
Two of those are forecasts and two are counts, and that distinction matters: only the robots are equipment that shipped and was paid for. But take the narrowest reading — the devices already in use and the robots already working — and the installed base is thousands of times larger than the AI fleet inside it.
The silicon inside each device is still trivial
Gartner puts the semiconductor content of an average car at $860 in 2024, rising to $1,902 by 2034. That is one class of device. A fridge, a shopping cart, a cargo trailer, a boat: each needs only a few dollars of silicon to become a customer of machine intelligence, and each is replaced on a cycle that has nothing to do with data-centre spending.
That is the asymmetry the bubble argument misses. Four years have gone into building the factory that makes a prediction cheap. Barely any of that cheapness has been put inside the objects people already buy every year.
The constraint is also the size of the opportunity
The strongest bear argument is electricity. Data centres used about 415 terawatt-hours in 2024 — a terawatt-hour being a billion kilowatt-hours, roughly the annual use of 90,000 European homes — and the International Energy Agency expects about 945 terawatt-hours by 2030.
That doubling is the bear case, and it is also the smallest number in the file: 945 terawatt-hours is around 3% of the world's electricity. Read it as scarcity and AI cannot grow. Read it as a share of the grid, and an industry has attracted this much capital while using three percent of the world's power. Mature industries do not look like that.
What would prove the bears right
- The four largest cloud buyers plan $720–740bn of spending in 2026. If that is guided down by more than 20% for 2027, the air pocket is real. - Nvidia's gross margin below 68% for two quarters running would say the equipment is being discounted rather than ramped. - AI-capable PCs stalling below 40% of shipments would say the edge is a marketing term rather than a product cycle. - Robot installations falling back below 600,000 a year would say the physical buildout is a subsidy story.
None of those are in the current data. But the bulls carry a burden of proof too, and the margin line is where it gets tested first.
The point is not that AI is big
It is that the cost of a prediction keeps falling while the number of objects that could afford one is 21bn and rising. The data centre is the factory. The devices are the market. The factory has been built and priced; the market has not.
That is what an infancy looks like: enormous capital, visible strain, and customers who have not arrived yet because the product is only now becoming cheap enough to fit inside them.