Castles in the Air: The Potential Financial Bubble in Artificial Intelligence

From railways to data centers, history shows how technological revolutions can create wealth for society without necessarily rewarding their investors.

AI-Talks.org | October 30, 2026 | 15 min read

Key Takeaway: Artificial intelligence may profoundly transform the economy and still generate financial losses. The risk does not necessarily lie in the technology itself, but in excessive valuations, mounting debt, and the possibility that future profits may fail to justify the trillions of dollars invested.

One of my favorite books on financial markets is A Random Walk Down Wall Street, by Burton G. Malkiel. First published in 1973, the book reached its 13th edition in 2023, marking its 50th anniversary. It contains a warning that remains remarkably relevant today:

“We should have known that investments in transforming technologies have often proved unrewarding for investors.”

Artificial intelligence may be the most striking example of this paradox in our time. As companies pour hundreds of billions of dollars into processors, models, and data centers, one question remains unanswered: who will manage to turn all this infrastructure into enough profit to justify its cost?

Malkiel revisits an idea associated with John Maynard Keynes, known as the Castle-in-the-Air Theory. Rather than valuing a company solely on its ability to generate future profits, investors may buy its shares simply because they believe someone else will be willing to pay even more for them. Prices then come to depend on expectations about the expectations of other market participants.

That is how castles in the air are built.

This mechanism does not require the technology itself to be an illusion. Quite the opposite: the more compelling the promise of economic transformation, the more willing investors may be to anticipate financial returns that could take decades to materialize. The trouble begins when enthusiasm for the future translates into valuations that leave little room for outcomes that fall short of expectations.

When Technological Revolutions Create Financial Bubbles

In Britain during the 1840s, the expansion of the railway network triggered one of the greatest waves of financial speculation of the nineteenth century. During the so-called Railway Mania, investors financed hundreds of projects, drawn by the prospect of participating in a revolution in transportation. Many railway lines, however, were planned without sufficient demand to justify their costs. As enthusiasm began to fade in 1845, share prices fell, companies collapsed, and investors lost fortunes. The railways themselves, however, remained, continuing to transform the British economy over the following decades.

More than a century later, during the expansion of the internet in the late 1990s, investor enthusiasm produced two interconnected phenomena: the speculative rise in technology stock valuations and excessive investment in telecommunications infrastructure. Anticipating extraordinary growth in internet traffic, telecommunications companies invested billions of dollars in fiber-optic networks. Demand did grow, but not quickly enough to generate adequate returns on all the capacity installed. Many companies went bankrupt or were restructured, while their networks remained in use and helped support the emerging digital economy.

Both episodes reveal the same economic mechanism: a genuine technological transformation generates expectations of growth, attracts investment ahead of future demand, and can ultimately produce excess capacity. When returns fall short of expectations, investors suffer losses, but much of the infrastructure they financed remains available to society.

Economist Carlota Perez, in Technological Revolutions and Financial Capital (2002), interprets this pattern as part of the major cycles of technological transformation. Initially, financial capital accelerates the deployment of new infrastructure, often accompanied by speculation and excessive investment. Subsequent corrections can eliminate less efficient companies, redistribute assets, and create the conditions for the technology to spread more widely and productively throughout the economy.

This is precisely the risk I see in the current expansion of artificial intelligence. The argument is not that AI is an empty promise, but that the scale and pace of investment in its infrastructure may be running ahead of demand and profitability that could take much longer than expected to materialize. In such a scenario, the technology would continue to advance and find new applications, even as some companies and investors failed to recover the capital they had committed.

The historical comparison, however, requires some important distinctions. Like railways and fiber-optic networks, data centers require enormous upfront investments whose recovery depends on future demand. Yet while railway tracks and cables have relatively long physical lifespans, specialized processors can lose their economic competitiveness within just a few years as newer hardware delivers greater computing performance per unit of energy and per dollar invested.

There is also another important difference: the efficiency of AI infrastructure can improve without replacing the hardware itself. One example is quantization, a technique that reduces the numerical precision used to represent the parameters of neural networks, often from 16 bits to 8 or even 4 bits. This reduces memory requirements and can accelerate model execution, depending on the hardware, although potential losses in output quality must be evaluated. Other techniques make it possible to process multiple requests simultaneously or reuse previously computed results, increasing the productivity of existing processors.

These advances make the profitability of AI infrastructure particularly difficult to predict. Greater efficiency can reduce costs, extend the useful economic life of equipment, and encourage new applications. Competition, however, may pass some of these savings on to customers, driving down the prices of computing services. If demand growth fails to offset this downward pressure on prices, or if newer processors make existing hardware less competitive, some investments may generate returns below expectations.

The historical parallel, therefore, does not lie in the physical similarities between railways, cables, and processors, but in the difficulty of anticipating how much capital will be needed, how quickly demand will materialize, and who will ultimately capture the profits generated by the transformation.

History thus reveals a paradox: investors can, to a certain extent, correctly anticipate a technological revolution and still lose money financing it. The problem does not necessarily lie in their predictions about the future, but in the price they pay to participate in it.

Amara’s Law and the Timing of Financial Markets

The relationship between financial expectations and technological change can also be understood through Amara’s Law, which I discussed in my article Amara’s Law and AI: Hype Today, Transformation Tomorrow. The well-known observation attributed to futurist Roy Amara suggests that we tend to overestimate the effects of a technology in the short term and underestimate them in the long term. It is not a mathematical law, but an observation about our difficulty in anticipating the pace of technological change.

In financial markets, this mismatch can have significant consequences. When a revolutionary technology emerges, investors try to anticipate its economic benefits. Expectations of growth drive share prices higher, attract capital, and encourage the expansion of infrastructure. However, investments must be made today, while the anticipated profits depend on future demand that has yet to materialize.

Artificial intelligence presents precisely this problem. Building data centers, purchasing processors, and training models require immediate expenditures of billions of dollars. Turning those capabilities into sustained productivity gains, however, depends on slower processes: technological integration, organizational restructuring, workforce training, application development, and consumer adoption. A company may demonstrate impressive technical advances without being able to translate them, at the same pace, into financial results proportional to the investments made.

This gap creates a tension between investors’ time horizons and the time required for a technology to mature. Financial markets anticipate revenues and profits that may not materialize for many years, while equipment depreciates, debts come due, and competition drives prices down. If the expected returns take longer to emerge, valuations may undergo substantial corrections, even as the technology continues to improve and find new applications.

The process can also work in the opposite direction. After a period of disappointment and declining investment, technologies that were initially overhyped may become more efficient, affordable, and widely adopted. This is what happened with the internet, whose economic and social consequences ultimately exceeded many of the expectations formed during the dot-com bubble, even though numerous companies from that era disappeared.

Amara’s Law cannot predict when a financial bubble will form or burst. Its value lies in revealing that the pace of innovation, the pace of economic adoption, and the pace of financial markets do not necessarily coincide. A technology may fulfill its promises over the long term while investors who anticipated returns too aggressively suffer losses in the present.

The possibility of an AI financial bubble, therefore, does not contradict the prospect of a genuine technological revolution. Quite the opposite: both can coexist. We may be overestimating the profitability of artificial intelligence over the next few years while simultaneously underestimating the depth of the transformations it could bring about over the coming decades.

Trillions of Dollars in Search of Returns

Initial public offerings (IPOs) often accompany periods of technological enthusiasm. During the dot-com bubble, between 1999 and 2000, IPOs in the United States generated average first-day stock price gains of approximately 65%, according to a study by Tim Loughran and Jay Ritter. Between 1990 and 1998, that average had been around 15%.

These figures reveal the intensity of demand for newly issued shares, although they also reflect how IPOs were priced. Investors who acquired shares at the offering price were in a very different position from those who bought them after the initial surge. For many in the latter group, the expectation of becoming wealthy ended in substantial losses.

Today’s artificial intelligence market includes companies that are financially far stronger than many of their dot-com predecessors.

On October 26, 2026, Anthropic, the developer of Claude, filed its IPO registration statement with the U.S. Securities and Exchange Commission (SEC). During the first half of the year, the company reported revenues of $16.3 billion and generated approximately $3.9 billion in operating cash flow, despite recording an operating loss of $693 million.

These results demonstrate a substantial commercial operation, not a company sustained by promises alone. Nevertheless, its growth involves long-term commitments to computing capacity and continued investment in research and infrastructure. Future profitability will depend on its ability to reduce the costs of operating its models without competition driving customer prices down even further.

The IPO could finance further expansion and provide liquidity to existing investors. For those who purchase its shares, however, returns will depend fundamentally on the valuation assigned to the company.

The financial expansion of artificial intelligence extends far beyond IPOs. According to J.P. Morgan projections, global capital expenditures related to AI could reach $900 billion in 2026 alone, with cumulative spending estimated at $5.5 trillion by 2030.

Much of this capital is being directed toward building data centers, purchasing processors, expanding communications networks, and developing energy infrastructure.

The International Energy Agency estimates that global electricity consumption by data centers could nearly double between 2025 and 2030, rising from 485 to 950 terawatt-hours (TWh) annually. Artificial intelligence is often presented as a digital technology, but its expansion depends on a continuous supply of electricity, transformers, transmission lines, and cooling systems. Energy availability is therefore not merely an environmental concern, but also a decisive factor in the cost and economic viability of AI investments.

Nvidia has emerged as one of the main beneficiaries of this race. In the second quarter of its 2027 fiscal year, which ended in July 2026, the company reported revenues of $96.2 billion, an increase of 106% over the same period a year earlier. Of that amount, $89 billion came from its data center business.

In financial markets, Nvidia shares had gained approximately 14.6% since the beginning of 2026 through October 29, when the company’s market capitalization stood at around $5.2 trillion. These figures demonstrate both the extraordinary expansion of its business and the exceptionally high growth expectations already reflected in its valuation.

These are extraordinary results. Yet processor manufacturers and their customers occupy fundamentally different economic positions. While Nvidia generates revenue by selling hardware, its customers must recover those investments by providing services that generate sufficient cash flow over time.

This creates an important asymmetry. The expansion may be profitable for infrastructure suppliers even if some of the companies operating that infrastructure fail to earn adequate returns on the capital they have invested.

Who Wins, Who Loses, and Who Ends Up Owning the Assets

One fundamental difference between the current expansion of artificial intelligence and the dot-com bubble lies in the financial strength of its major participants.

Microsoft, Alphabet, Amazon, and Meta have established businesses, diversified revenue streams, and extensive access to capital markets. Microsoft, for example, generated approximately $183 billion in operating cash flow during its 2026 fiscal year, while its expenditures on property and equipment reached around $116 billion.

These companies also have strategic reasons to continue investing. Artificial intelligence could reshape entire markets, including search, advertising, software, and cloud computing. Falling behind in this transformation could threaten the very businesses that currently sustain their profitability.

Investing in AI therefore serves as a form of competitive protection. However, decisions that are individually rational can produce collectively inefficient economic outcomes. If several companies build infrastructure to compete for the same market, the resulting capacity may exceed the level of demand needed to generate adequate returns. For the major technology platforms, the most likely consequences would be narrower profit margins, reduced free cash flow, and declining share prices, rather than necessarily a solvency crisis.

The situation is different for specialized infrastructure operators, often referred to as neoclouds. These companies frequently rely on loans, long-term contracts, and equipment-backed financing to sustain their expansion. CoreWeave, backed by Nvidia and publicly traded since 2025, has become an example of this vulnerability. According to Reuters, its debt was approaching $50 billion in October 2026. Despite holding substantial contracts with major technology companies, it must continually finance infrastructure that requires enormous upfront investments.

Another troubling aspect is circular financing, in which suppliers, customers, and investors maintain financial relationships that reinforce one another. Imagine a processor manufacturer investing in an AI company. That company uses part of the funding to purchase services from a data center operator, which, in turn, buys equipment from the same processor manufacturer. Capital flows through different companies within the ecosystem, financing real economic activity and generating commercial revenues, but also creating financial interdependence among its participants.

Such arrangements are not necessarily improper or artificial. They can accelerate technological expansion and make investments possible that would otherwise be difficult to finance. The risk emerges when a company’s revenue growth depends, in part, on the funding that it or its partners are providing to customers. If end-user demand falls short of expectations, financial difficulties can spread throughout this network of contracts and financing arrangements.

In October 2026, the International Monetary Fund warned about these mechanisms. An analysis of 14 transactions undertaken since 2023 identified approximately $300 billion in equity investments and $650 billion in debt, associated with $1.4 trillion in future revenue commitments. These figures do not represent inevitable losses, but they illustrate the scale of the financial obligations sustained by expectations of continued growth in the sector.

If overinvestment occurs, the most heavily indebted operators may be the first to encounter difficulties. Their equipment and facilities, however, will not simply disappear. They may be sold, refinanced, or acquired by financially stronger companies. In such a scenario, a crisis for some investors could become an opportunity for their competitors, who might acquire depreciated infrastructure or secure computing capacity at prices below those originally anticipated.

Some of the losses could fall on shareholders, creditors, and private credit funds. Although mechanisms exist to absorb such losses, the concentration of debt, collateral, and contracts among financially interconnected participants also creates risks of financial contagion. The extent of these effects will depend on credit conditions and the ability of the companies involved to renegotiate their obligations.

The infrastructure may survive a financial correction, but its ownership, its value, and the distribution of the profits it generates could change profoundly.

The Paradox of Technological Abundance

Artificial intelligence is already delivering tangible results in programming, scientific research, medicine, education, and a wide range of business activities. Yet adopting AI requires more than purchasing access to models and installing systems. Organizations often need to redesign their processes, train their employees, and develop new approaches to management.

Erik Brynjolfsson, Daniel Rock, and Chad Syverson describe this phenomenon through what they call the Productivity J-Curve. General-purpose technologies may require years of complementary investments before their benefits become fully visible in productivity statistics.

A Gartner survey released in October 2026 found that only 13% of the technology executives surveyed reported that their organizations had achieved significant or transformative value from AI.

This finding does not demonstrate that the technology has failed. It may reflect the time required to reorganize business processes and translate technical capabilities into economic gains. Moreover, AI does not need to generate direct revenue to create value. It can reduce operating costs, accelerate product development, and improve existing services.

Perhaps the most interesting question, however, is how these benefits will be distributed.

As AI models become more efficient and affordable, their potential applications expand. Lower costs can encourage new uses and increase the size of the market. At the same time, competition can drive down the prices charged by providers, putting pressure on their profit margins.

In other words, artificial intelligence may create an enormous economic surplus without the companies building and operating its infrastructure necessarily capturing a proportional share of that value.

Society could therefore enjoy extraordinary productivity gains while certain investors experience disappointing returns. Excess computing capacity could make AI services cheaper and accelerate the spread of the technology, even after financial losses and corporate restructuring.

None of this means that a bubble is necessarily about to burst. Demand continues to grow, major suppliers are reporting impressive results, and new applications are constantly emerging. Some of today’s valuations may ultimately be justified by future growth.

But technological prosperity does not eliminate the possibility of overinvestment. The history of railways and telecommunications demonstrates that the benefits of a technological revolution can endure long after the financial expectations of its early investors have proved excessive.

The central question is not when a crisis might occur, but who has assumed the risks, how much capital will be needed to sustain the expansion, and who will ultimately capture the economic benefits it creates.

Malkiel’s warning remains relevant because it addresses a recurring mistake: confusing the importance of an innovation with the financial value of any company associated with it.

We may be entirely right about the future of artificial intelligence and profoundly wrong about how much we are willing to pay for that future.

---

References

1. Alphabet Inc. “Alphabet Announces Third Quarter 2026 Results.” October 28, 2026.
2. Alphabet Inc. “Q3 2026 Earnings Call Transcript.” October 28, 2026.
3. Anthropic PBC. Form S-1 Registration Statement. U.S. Securities and Exchange Commission, October 26, 2026.
4. Brynjolfsson, Erik, Daniel Rock, and Chad Syverson. “The Productivity J-Curve: How Intangibles Complement General Purpose Technologies.” American Economic Journal: Macroeconomics 13, no. 1 (2021): 333–372.
5. Chapekis, Athena, and Anna Lieb. “Google Users Are Less Likely to Click on Links When an AI Summary Appears in the Results.” Pew Research Center, July 22, 2025.
6. Federal Reserve Bank of St. Louis. “NASDAQ Composite (NASDAQCOM).” FRED, Federal Reserve Economic Data. Data provided by Nasdaq, Inc.
7. Garber, Peter M. “Famous First Bubbles.” Journal of Economic Perspectives 4, no. 2 (1990): 35–54.
8. Gartner. “Gartner Survey Finds Only 13% of CIOs Say Their Enterprise Has Realized Significant or Transformational Value from AI.” October 19, 2026.
9. Goldgar, Anne. Tulipmania: Money, Honor, and Knowledge in the Dutch Golden Age. Chicago: University of Chicago Press, 2007.
10. International Energy Agency. Key Questions on Energy and AI. Paris: International Energy Agency, 2026.
11. International Monetary Fund. Global Financial Stability Report: Resilience, Risk, and Transformation. Washington, DC: International Monetary Fund, October 2026.
12. J.P. Morgan Global Research. “What Unprecedented AI Capex Means for Markets and Investors.” October 19, 2026.
13. Loughran, Tim, and Jay R. Ritter. “Why Has IPO Underpricing Changed Over Time?” Financial Management 33, no. 3 (2004): 5–37.
14. Malkiel, Burton G. A Random Walk Down Wall Street: The Time-Tested Strategy for Successful Investing. 50th Anniversary Edition. New York: W. W. Norton & Company, 2023.
15. Microsoft Corporation. Microsoft 2026 Annual Report. Redmond, WA: Microsoft Corporation, 2026.
16. NVIDIA. “NVIDIA Announces Financial Results for Second Quarter Fiscal 2027.” August 26, 2026.
17. Perez, Carlota. Technological Revolutions and Financial Capital: The Dynamics of Bubbles and Golden Ages. Cheltenham: Edward Elgar, 2002.
18. Pinheiro, Maurício Veloso Brant. “Por que superestimamos os próximos dois anos e subestimamos os próximos vinte.” AI-Talks.org, July 11, 2026.
19. Pinheiro, Maurício Veloso Brant, and Ricardo Brant Pinheiro. “The Rise of Artificial Intelligence, The Looming Energy Crisis and The Future of Nuclear Power.” AI-Talks.org, November 18, 2024.
20. Reuters. “CoreWeave’s AI Contracts and the Lenders Financing Its Build-Out.” October 15, 2026.
21. Reuters. “Wall Street Closes Sharply Lower as AI Cash Burn Fears Spark Tech Selloff.” October 30, 2026.
22. StocksToBuyNow.ai. “Stock Market Heatmap: Daily Close, October 30, 2026.” October 30, 2026.

---

Maurício Veloso Brant Pinheiro, PhD
Professor of Physics, Federal University of Minas Gerais (UFMG)
Founder, Author and Editor, AI-Talks.org
About Us

Editorial transparency note: This article, as with all articles published on this site, was conceived, directed, written, and reviewed by Prof. Maurício Veloso Brant Pinheiro. Artificial intelligence was used as an assistant for editorial refinement, formatting, image generation, SEO metadata, and publication workflow.

---

Copyright 2026 AI-Talks.org