Here is a funny thing about valuation “bubbles” involving financial instruments. In every individual case, you unavoidably end up asking: “What were they thinking?” Or stated differently, what causes a significant part of the supposedly rational investing public to collectively lose their minds for a relatively brief period? A few tulip bulbs, for example, were valued more highly than a modest Amsterdam house in 1637, until they weren’t. Leveraged speculation in frontier commodity markets of Asia and America led to the South Sea Bubble and the Mississippi Bubble in Europe in the 1720s.
The seventeenth and eighteenth-century bubbles were commodity bubbles: tulips, spices, or beaver pelts. But the Railway Mania in the UK in the mid-1840s focused on an amazing new technology, railroads, marketed to the public as risk-free investments. That the British railroad companies’ shares ultimately soared and then lost more than two-thirds of their value, immiserating a large portion of the investing public, is almost besides the point. Bubbles always end badly by definition. But the UK’s railway mania is instructive because it was the first truly modern bubble which had all three components of more recent financial bubbles: 1) a new technology, railroads for moving goods and people, the long term economic impacts of which are unclear; 2) financial innovation usually involving leverage or speculation and 3) a government that either participates in the scheme or has little inclination to intervene in financial markets (until it’s too late).
Related: Tanker Explosion Off Yemen Deepens Red Sea Security Fears for LPG Trade
Using this as a framework, it seems fair to ask if today’s AI stocks, which have substantially outperformed broad market indices, are in a financial bubble? But we’re utility analysts asking this question for a simple reason. The demands for new electric power-generating facilities supposedly required by AI are simply staggering both in amount and expense. For example, the highly respected Electric Power Research Institute (EPRI) expects new AI-related electricity demand to reach 50 gigawatts (GW) by 2030. A Deloitte study showed AI electric demand increasing from 4GW now to 123 GW by 2035. And Sam Altman’s Open AI Consortium (including Softbank and Oracle) stated it would need roughly 250 GW by 2033, which at today’s prices would cost over half a trillion bucks to build. (To put these numbers in context, total generating capacity in the USA, presently, is 1,300 GW, growing about 30 GW per year.) All we can say is that with respect to power costs, the AI industry is behaving as if they are either irrelevant or simply a modest expense item, neither of which is true.
Utility assets and AI data centers resemble each other in that both are highly capital-intensive, but with one big difference. The heart of a utility is a power plant that lasts forty years (an annual depreciation rate of 2.5%). The primary cost component of a data center is the vast array of computer chips, which have a relatively short life, maybe 2-3 years. A capital asset with a two-year life has a 50% depreciation rate, creating a very large expense item on the income statement, which has to be covered by revenues from the business or by increased corporate borrowing. That still doesn’t make the AI/data center boom a bubble, but it does suggest our next concern.
Capital expenditures for AI and data centers seem high. Not just to us. In fact, economists have noted that in the first half of 2025, AI-related capital expenditures contributed more to US domestic growth than consumer spending. This marks a dramatic reversal of prior trends, where consumer spending has been responsible for the bulk of economic activity. All we can conclude here is that this highlights the biggest financial risk for AI investors: will all these new facilities earn a competitive return on all this new, invested capital? Will there be enough revenues to support these new assets? We know the level of planned capex and that there are exciting uses for this technology. We do not know how much people will pay for these services, and therein lies a huge risk.
This boom (bubble?) reminds us of the power-generating bubble that burst in the early days of the century. When the power generating market was opened to competition, a host of legacy electricity firms, bankers, entrepreneurs, pipelines, European utilities, and crooks piled in to build a multitude of power plants, encouraged by favorable forward price curves and plenty of low-cost borrowed money. They knew that all the construction (if completed) would produce a power glut, but they all believed that the other guy would back down and not build, thereby preventing an oversupply. Well, the other guys did not back down, demand for power was weak, and carnage followed.
There are other related risks here, such as the role of vendor financing. The biggest players in this area (Microsoft, Google, Meta, and Nvidia) are financially robust. But they support an ecosystem of smaller vendors and suppliers who rely exclusively for their infrastructure and software on the aforementioned giants. This creates a circular financing scheme where one group sells products to another wholly captive or financially dependent group while purchasing their services. In a bubble, this is all financed with common stock issuances at ever higher valuations. Which works until it doesn’t. Then the house of cards falls down, as it did in the housing and dot-com bubbles earlier this century.
The AI boom assumes that the tech bros in California, choosing an AI strategy, picked the right horse. The Chinese seem to think otherwise, offering less expensive AI solutions to less expensive problems, more or less. Chinese industry has shown, with government help, a remarkable ability to take ideas from abroad and commercialize them into a lower-cost product. AI is in their targets. That may put these huge American AI investments at risk. So don’t start to count the profits.
Let’s conclude with these observations about financial bubbles. First, there is a kind of conflation in the term between genuinely nonsensical ideas and those viable corporations whose shares were simply grossly misvalued. When the dot-com bubble burst around the year 2000, the shares of Amazon declined by over 90%. But it is still around, bigger than ever. Pet food supplier Pets.com, on the other hand, which lost money on every sale, is not. If we go back to our original example of the British railway mania of the 1840s, we can observe something similar. Those railroads, which provided an enormous economic boost to a rapidly industrializing British economy, also caused grievous financial injury to a generation of investors because of the wild misvaluation of the fledgling industry employing a radically new technology, the steam engine. But the legacy of that technology bubble was a fully functioning national railway system that still exists. So, after this AI and data center investor euphoria passes, the investors lose their money and the smoke lifts, what remains, just ashes or many valuable, profitable, important businesses?
By Leonard Hyman and William Tilles for Oilprice.com
More Top Reads From Oilprice.com:
