Our thesis: the AI data-center build-out is real, and it is also a capital-spending cycle. Cycles reward the sellers of equipment on the way up and punish them on the way down, often well before the technology itself stops mattering. So the question we ask is not “is AI important?” It clearly is. The question is “where are we in the spending cycle, and who carries the risk if buyers pause?”
A story we have seen before
Picture a new technology that everyone agrees will change the economy. Companies race to build the infrastructure for it, because being late feels more dangerous than overspending. Suppliers of the picks and shovels report record orders. Investors extrapolate those orders into the future. Then, at some point, buyers realise they have built more capacity than today’s customers can use. They slow down to digest what they own. Orders for suppliers fall, sometimes sharply, even though the technology keeps spreading.
That story has played out several times. It does not tell us the AI build-out will end the same way. It tells us which questions to ask.
Example 1: Britain’s railway mania
In the 1840s, British investors poured money into railway companies. Railways really did transform trade and travel. But far more lines were proposed and financed than the country could profitably use, many projects were never completed, and a large share of investors lost money when the boom collapsed. The lesson we take: a technology can be revolutionary and still be over-built in the short run.
Example 2: fiber optics and the dot-com bust
In the late 1990s, telecom companies laid enormous amounts of fiber-optic cable, betting that internet traffic would keep exploding. Traffic did grow, for decades. But the build-out ran ahead of demand. When the dot-com bubble burst in 2000, many telecom carriers went bankrupt, and much of the fiber sat unused (“dark fiber”) for years. Equipment makers that had sold the gear saw orders collapse. Cisco, which briefly ranked among the most valuable companies in the world in 2000, saw its stock fall steeply and took many years to recover, even as the internet it helped build became part of everyday life.
The lesson: the internet won, and many of the people who paid for its first build-out still lost. Value later flowed to companies that used the cheap capacity, not only to those who built it.
Example 3: Nvidia’s own earlier cycles
Nvidia started as a maker of graphics chips for video games. In 2006 it introduced CUDA, software that let developers use its chips for general computing, which later made them the standard tool for training AI models. Its path was not smooth. Demand from cryptocurrency mining boosted sales and then faded, and in 2018 the company was left with excess inventory of gaming chips that weighed on results. After ChatGPT launched in late 2022, demand for data-center chips surged and Nvidia became one of the largest companies in the world.
The lesson: even the best-positioned supplier has lived through boom-and-digest phases. Leadership in a technology does not remove cyclicality from orders.
What is different this time
To be fair to the bull case, today’s biggest buyers are not fragile start-ups funded by junk bonds, as many 1990s telecom carriers were. Several are among the most profitable companies ever, and they can fund spending from cash flow. AI services are also already being sold to customers, not only promised. Those are real differences, and they are why we do not assume a repeat of 2000. They change the odds and the severity, not the basic shape of a capital cycle.
What we watch
- Buyers’ capital-expenditure guidance. Large cloud and internet companies discuss spending plans on earnings calls. The direction of the plans matters more than one quarter.
- Return on the spending. Listen for revenue from AI products, not only capacity added. Spending without matching revenue is the classic warning sign.
- Supplier backlog, lead times and inventory. Suppliers describe these in 10-Q and 10-K filings. Shrinking lead times or rising inventory can be early signs of cooling, as in 2018.
- Customer concentration. Filings disclose when a few customers make up a large share of revenue. More concentration means more risk if one buyer pauses.
- How the spending is financed. Spending from cash flow is sturdier than spending funded by rising debt. Debt-funded build-outs drove the worst of the telecom bust.
- Physical limits. Electricity, grid connections, land and cooling can slow a build-out regardless of demand.
Illustrative example, not real data.
Imagine a supplier where one customer is 40 of every 100 dollars of revenue. If that customer cuts orders by a quarter, the supplier loses 10 of every 100 dollars of sales before anyone else changes plans. The numbers are made up to show how concentration amplifies a pause.
How we see the future
Bull scenario. AI products earn enough revenue that buyers keep expanding, efficiency gains create new uses rather than less demand, and physical limits ease. Suppliers keep growing, with only mild pauses.
Base scenario (our view). Spending keeps rising for a while, then enters a digestion phase where buyers absorb capacity. Supplier results become bumpier than investors expect, while the technology keeps spreading. Over time, more of the value shifts toward companies that use AI well.
Bear scenario. Revenue from AI lags far behind spending, buyers cut budgets at the same time, and debt-funded projects run into trouble. Suppliers face a sharp order drop, echoing the fiber bust, even though AI remains important.
What would change our mind. We would lean bullish if buyers report AI revenue growing in step with spending for several quarters. We would lean bearish if several large buyers cut spending guidance together, if supplier inventories climb, or if build-outs increasingly rely on borrowing.
Our take
We treat the AI build-out as a cycle to understand, not a straight line to extrapolate. History says the technology can win while the first wave of spending disappoints. Read what buyers say about spending, read what suppliers say about orders, and pay attention when those two stories start to disagree. This is our opinion for education, not a recommendation about any stock.