True AI believers place their faith in the transformative powers of the technology, sound funding of the capital expenditure, and the ability of firms and investors to manage the associated risks. Unfortunately, “it ain’t what you don’t know that gets you into trouble, it’s what you know for sure that just ain’t so.”
First, technological progress alone does not determine investment success.
Analysis of past investment booms highlights that even if an innovation proves revolutionary, identifying the ultimate winning firms is difficult. The process may take longer than investors expect, and overcapacity and poor capital allocation destroy wealth in the meantime. Often, the ultimate beneficiaries are not the pioneers but later entrants able to acquire distressed assets cheaply.
AI economics are questionable. Given the expected $8-10 trillion investment, one study found that by 2032 annual revenues of $3.7 trillion, (over 9% of estimated US GDP), would be needed to generate a modest unlevered return of 10%. This assumes a 50% cash flow margin and a generous six-year life for the chips. Notwithstanding forecast Total Addressable Markets of $30 trillion (effectively encompassing the global economy), the required increases may not be achievable.
Corporate users are increasingly seeking to limit AI spending or demonstrate clear returns before expanding usage. The majority of individual users are currently non-paying and transitioning them to subscription models is difficult.
Margins are already declining because of the falling price of tokens, by 99% in some cases, and rising expenses for electricity and cooling water, data acquisition to train models and compliance.
Given growing global community resistance to data centres, construction delays will create shortages and increase cost pressures. Firms like Anthropic and OpenAI may struggle to maintain performance advantages and pricing power in the face of growing competition from cheaper Chinese suppliers and open-source models.
These dynamics suggest many AI ventures may find it difficult to meet revenue, earnings, cash flows and return on investment expectations.
Second, investors may be underestimating the risks embedded in the funding structure supporting the AI boom.
The central assumptions are that it is supported by real risk equity capital and that borrowings are underwritten by large AI investors, the hyperscalers - Amazon, Alphabet, Apple, Meta, Microsoft and Oracle as well as chip makers like Nvidia and Broadcom - which have profitable core businesses, strong earnings and strong balance sheets.
In practice, much of this capital may be less patient or permanent than it appears. Venture and private-equity funds have finite lives and ultimately require exits through trade sales or public listings. As a result, continued access to new capital remains important as many ventures will have negative cash flows and need funding for some time.
The funding from hyperscalers and chipmakers was initially from free cash flow. Prior to the present boom, hyperscalers generated free cash flow of around $70 billion. This has reversed to a shortfall of $1.6 billion as AI investments absorb available funds, requiring borrowings and eventually new equity issues.
Current borrowings are running at around $500 billion per annum. AI related ventures have outstanding debt of around $1.2 trillion (14% of all investment-grade debt) which is increasing rapidly. There is growing reliance on credit from private credit funds.
Alphabet, Microsoft, Amazon, Meta and Oracle also have as much as $1.65 trillion of off-balance obligations, compared to $1.35 trillion in traditional debt.
One study found that these commitments lift indebtedness levels by nearly 150% on average and materially weakens their credit quality. There are additional guarantees, revenue support agreements and underwriting of the residual value of processing chips, whose impact is difficult to estimate due to limited disclosure.
A complex ecosystem of AI infrastructure providers and data-centre operators has emerged. Outside of the stronger hyperscalers and chipmakers, many rely on private credit and other forms of debt financing that are typically more expensive and shorter-term than traditional corporate borrowings. Within this highly interconnected system, all participants are totally dependent on the same underlying earnings and cash flows. If AI revenues fail to meet expectations, the inherent circularities mean that financial stress will spread rapidly beyond the model developers themselves, to the hyperscalers, equipment suppliers and infrastructure providers and ultimately to lenders and investors.
Finally, investment in generic infrastructure does not protect capital. Cisco, for which the investment case in the 1990s was similar, briefly became the world’s most valuable company on the assumption that its routers and other products would be crucial to the Internet irrespective of the industry’s evolution.
While the company’s performance has been steady, investors in Cisco lost out as its share price plummeted in 2000 only reaching the same level after 25 years.
One reason was excess capacity. During the telecoms and fibre-optic boom, investors drastically overestimated demand. Much of the capacity built during that period remains underutilised today. Only around 50% of installed fibre-optic capacity is lit, while average global network utilisation is approximately 26%. The current expansion may not avoid this trap.
AI investment is 17 times that of the 2000 dot com boom and four times the 2008 sub-prime housing bubble. When the dot com boom ended, even survivors like Microsoft, Apple, Oracle and Amazon fell 65, 80, 88, and 94% respectively taking 16, 5, 14 and 7 years to recover their 2000 peaks.
Just as trees do not grow to touch the sky, investment bubbles inevitably deflate when returns fail to live up to expectations and financing dries up.
Notes: All amounts are in US dollars.
Satyajit Das is an internationally respected expert in finance, who presciently anticipated the 2008 global financial crisis. In 2014 Bloomberg nominated him as one of the 50 most influential financial thinkers in the world. He is an author, whose latest book The Everything Bubble (Monash University Press) is out in 2027.
© 2026 Satyajit Das All Rights Reserved