When DeepSeek released its low-cost large language model in January 2025, semiconductor shares fell. The logic was simple. If a model can be built cheaply, the expensive GPUs matter less; efficiency rises, so demand falls.
That logic has, so far, been wrong.
The unit price of inference has dropped by nine-tenths since 2023. Yet the tokens consumed have not fallen but exploded. On Bain & Company’s reckoning, the price per token halved between December 2024 and December 2025 while tokens consumed grew 4.5 times over the same period. Microsoft’s chief executive, Satya Nadella, greeted DeepSeek’s release with the line “Jevons paradox strikes again”. The nineteenth-century economist William Stanley Jevons found that as the steam engine grew more efficient, coal consumption rose rather than fell. A resource that gets cheaper is not used less; it opens new uses and its total climbs.
So the question “is AI real, or is it a bubble” is the wrong one. That demand is real is no longer the point at issue. The point lies elsewhere.
＊
The question to ask is the distance between spending and payback.
The money now going into AI infrastructure is not a paper hope. TSMC spent $26.8bn on plant in the first half of 2026. Mitsubishi Heavy’s gas-turbine business holds ¥2.2trn in customer prepayments. Hitachi Energy’s order backlog is $57.9bn, turning into revenue over as long as six years. Money spent, money received, money contracted. The structure of each is set out elsewhere.
Here is the decisive difference from the dotcom era of 2000. Then, monetisation came later; companies with no revenue carried a valuation. Now the reverse holds: the money moves first, and it moves in earnest.
But money moving does not mean money coming back. The spending is real. The payback is not yet tested. It is in the gap between the two that the lesson of a quarter-century ago sits.
＊
Cisco Systems in 2001 is worth recalling.
In April 2001 Cisco wrote down $2.25bn of inventory. It was among the largest such write-downs on record, and it tipped the company to its first loss in 11 years, on a quarterly net loss of $2.69bn. In the quarterly report it filed with the Securities and Exchange Commission, Cisco attributed the write-down to a sudden and significant fall in forecast revenue. The inventory was measured against a benchmark of more than 12 months of demand for each product.
Why hold such stock? The summer before, in 2000, Cisco’s orders were overflowing while its assembly lines stalled for want of parts. The company chose to build supply, committing to buy components months before they were needed. It lent its contract manufacturers $600m interest-free to buy parts on its behalf. Fearing a shortage, it committed ahead.
While demand held, this looked like a smart move made early. But the moment demand fell away in 2001, the stock beyond 12 months of need turned into a write-down with nowhere to go. Much of it was custom-built, and could not be resold.
Many accounts tell Cisco’s write-down as a story of double ordering: customers, fearing shortages, placed the same order with several suppliers, and apparent demand swelled. But the primary filing with the SEC does not go that far. What the company wrote is only that its own forecast had been too high. The lesson is not in the customers’ double ordering. It is that the supply side, fearing a shortage, committed to demand ahead of time.
＊
The same flow of money is running now.
Customers pay years before they take delivery of a turbine. That is what ¥2.2trn of prepayments is. TSMC’s management, at its 2026 results briefing, gave two reasons for raising capital spending: pressure from customers to build capacity, and buying tools ahead at inflation prices. In its fear-driven, forward commitment against a shortage, the structure is the same as Cisco’s in the summer of 2000.
There is one large difference. Cisco carried the inventory itself, and so took the write-down. Mitsubishi Heavy and Hitachi Energy are on the receiving end of prepayments; the inventory risk sits in a different place. If anything, it is the customer paying ahead.
Even so, the receiving end is not free of risk. A prepayment is a liability on the books, and for the years until delivery the cash is put to work as working capital while the obligation to deliver remains. The position of the risk differs; the risk does not vanish.
Which turns the question around. In this supply chain, who is first to hold inventory with nowhere to go?
＊
Three signals point to the same place.
TSMC’s management said, at the same briefing, that among mature process nodes only the AI-related ones are tight: the power-management ICs that data centres need in bulk, and sensors. These are made on old-generation technology, at 0.18 microns or 90 nanometres. Demand for consumer products, the company said plainly, is not strong.
The price of inference carries the same asymmetry. The unit price has fallen nine-tenths, but not evenly. Output from the top models still holds at $25-30 per million tokens. Commoditised open models have dropped to $0.09-0.18. That two-order gap has not closed. The frontier holds its price; only the commodity end sinks to the floor.
And Tokyo Electron’s operating profit for the year to March 2026 fell 10.4%. Demand for chipmaking equipment was not strong across the board. What was weak was the part aimed at general-purpose servers and PCs.
All three say the same thing. The economy is splitting along the seam between “AI-related” and everything else. What is tight is the frontier: high-bandwidth memory, the newest GPUs, power-management ICs. What sinks first is the commodity end, general-purpose servers and the older, general-purpose memory such as NAND. That is where the stock with nowhere to go would pile up first, if it piles up anywhere. The canary sits at the commodity end, not the frontier.
The share price does not do this selecting in a day. On a day when the market sells, frontier and commodity fall together, in order of beta. That is what the tape showed. But inventory and write-downs, unlike the share price, select over time. Recall how long it took Cisco’s $2.25bn to build.
＊
One last thing worth holding in view: what remains after the crash.
In the dotcom years the Telecommunications Act of 1996 set telecom companies pouring money into fibre-optic networks. Wavelength-division multiplexing sent the capacity a single fibre could carry leaping, and under the cry that laying it would bring the use, far more was laid than demand could meet. When it broke in 2001, more than 60 companies in the sector went bankrupt and the price of bandwidth collapsed. Most of the fibre laid stayed in the ground, unlit. The industry called it dark fibre.
But the fibre did not vanish. It stayed as physical plant. As broadband, video streaming and then cloud generated demand through the 2000s, that sleeping capacity was bought at distressed prices and became the ground the next era stood on.
Carlota Perez traces this pattern through the history of technological revolution. In the installation phase, speculative capital overbuilds the infrastructure. It crashes. But the overbuilt infrastructure remains, and on it the next real economy is deployed. Fibre was such a case.
The data centres, grids and transformers now being built will, if the bubble bursts, remain as physical plant too. Some of it, at least.
But for AI the pattern only half holds. Fibre, once laid, cost almost nothing to keep and could sleep 10 years and wait for the next wave of demand. Power plant, grids and shells are long-lived assets too. GPUs obsolesce in three to five years. In the years they wait, the value itself evaporates. Some physical plant remains; some obsolesces.
And here too the seam runs in the same place. What lasts is power, transmission and the buildings, the pick-and-shovel side. What obsolesces fast is the GPU, the most perishable of the tools for mining the gold. What dark fibre teaches is not that the physical remains. It is which physical remains, and which does not.
＊
The spending is real. Some of the physical plant will remain.
But which spending is recovered, and which plant survives into the next era, is not yet in the price. The word “AI” still bundles the side that is recovered with the side that turns into a write-down. The place to tell them apart is not the glamour of demand, the shipments of GPUs or the size of the backlog. It is at the end that cracks first, the canary at the commodity end. Demand may well keep climbing, as Jevons says. But that demand rises is one thing; that it is met at a price that pays the spending back is another.
A quarter-century ago Cisco’s orders were real. TSMC’s spending is real. The trouble is that neither, being real, ever promised the payback.
Nothing here is investment advice.