Jeff Bezos spoke recently at Italian Tech Week 2025 in Turin, where he distinguished between two types of bubbles.
The first is a financial one in which leverage creates a bubble that ultimately hurts the economy and crashes stock markets. The best example is the 2008/2009 global financial crisis, which was years in the making and had at its heart dodgy home loans packaged as A-grade debt. These famously included “ninja loans” — loans to people who had no income, no job and no assets. In other words, there was absolutely no way to verify their ability to repay the loan, and that ended exactly as you would expect.
The second is an industrial bubble, where new technology drives the hype. The bubble might burst, but at least it leaves behind useful infrastructure.
Former telecommunications company WorldCom is an example. In the late 1990s and early 2000s, at the time of the dot-com bubble, its revenues were declining. To hide this, its management committed fraud and the company collapsed. But its fibre-optic network remained valuable, and WorldCom was bought out of bankruptcy by Verizon in 2006.
Another example is the railroad building boom of the late 19th and early 20th centuries. The companies building the railroads overextended and most went bust, leaving their assets, the railroads, to be bought out of bankruptcy.
Buyers of such assets get them way below the cost of building, and in both examples, that infrastructure is still being used.
So the data centres under construction will be out of date in a few years and considered ancient within a decade
Bezos’s message is that the bubble in AI is an industrial bubble in that the data centres being built will have utility for decades to come, and while there may be significant overspend and capacity right now, society in general and big tech in particular will benefit. Further, he is not suggesting that the hyperscalers such as Alphabet, Microsoft, Amazon, Meta Platforms and the like will go bust, but we may see massive writedowns on the value of the data centres.
Broadly, his logic is solid, but it misses a key point.
Unlike railroads and fibre-optic networks, data centres age quickly. Nvidia is upgrading its chips about every 18 months, not only with a lot more computing power but also with less energy needs.
So the data centres under construction will be out of date in a few years and considered ancient within a decade. Sure, you can upgrade them, but that kind of defeats the point of his argument, as those upgrades won’t be cheap.
So this is bubble territory, but the difference between this and the dot-com bubble is that there is strong revenue, and real cash flow is being used to build the data centres, rather than debt in the previously mentioned infrastructure bubbles.
One caution: Meta is planning on using debt next year to fund new data centres. That is a real worry.
The writer holds shares in Nvidia
Would you like to comment on this article?
Sign up (it's quick and free) or sign in now.
Please read our Comment Policy before commenting.