The most persuasive infrastructure stories are often the ones that contain enough truth to make investors careless.
In June 2024, the Bank for International Settlements — the central bank of central banks — published its Annual Economic Report with an unusual set of warnings about artificial intelligence. The BIS was not skeptical of AI’s technological potential. It was specifically concerned about the financing structure gathering around AI infrastructure investment: the combination of high debt tolerance, compressed return timelines, and valuation assumptions that had become embedded in the capital cycle supporting data center construction, GPU acquisition, and cloud capacity expansion.
Railway Mania in the 1840s had shown how a genuinely transformative infrastructure could still produce a financing structure that eventually collapsed. The canals and railroads were real; the traffic they eventually carried was real; the economic value they created was substantial. But the capital that rushed into them faster than the cash flows could justify created a financing overhang that produced a severe correction when return expectations met slower-than-projected reality.
The Railway Mania comparison has become a standard reference for the AI infrastructure cycle. The BIS warning added something more specific: institutional analysis of how the credit and valuation assumptions around AI infrastructure specifically might create systemic risk, not merely speculative loss.
The World Before the Fortune
Transformative infrastructure cycles almost always attract more money than the earliest economics can fully justify. Railways, canals, telecom, and data systems all show the same structural tension: long-term usefulness does not automatically protect short-term financing from overreach.
The AI infrastructure cycle that began accelerating in 2022 had genuine technical foundations. Large language models demonstrated capabilities that earlier machine learning systems had not achieved. The computational requirements of training and running these models created real demand for GPU capacity, high-bandwidth memory, specialized networking, and the power infrastructure to support them. The hyperscaler capital expenditure plans announced by Microsoft, Google, Amazon, and Meta represented credible commitments to a technology platform with demonstrated commercial applications.
But the BIS report focused on what it called the “AI capex boom” from a financial stability perspective, not a technology assessment perspective. The concern was specific: a combination of private-credit vehicles funding speculative AI infrastructure, equity market valuations for AI-adjacent companies that implied return assumptions well above historical technology infrastructure norms, and corporate debt issuance by companies planning to invest heavily in AI infrastructure without fully demonstrated commercial return paths.
The environment favored rapid capital deployment because the competitive dynamics of the AI market rewarded early investment. Companies that built compute capacity early could develop models before rivals, attract engineering talent, and establish commercial relationships with enterprise customers. The competitive pressure to invest created a collective action problem: individual companies had incentives to deploy capital faster than the financial fundamentals of any individual project justified, because the alternative was falling behind competitors who were doing the same.
The Rise
The AI buildout has drawn attention because the opportunity looks historic. That very plausibility makes it easier for capital markets to normalize huge spend before the surrounding return assumptions are fully proven.
The BIS analysis identified several specific mechanisms of financial risk. First, private credit markets had expanded significantly into AI infrastructure financing — providing debt capital to data center developers, GPU lease originators, and smaller AI companies that could not access public debt markets. Private credit typically involves less transparency than public bond markets and may have embedded valuation assumptions that are difficult to verify until the underlying assets are tested by market conditions.
Second, the equity valuations of AI infrastructure companies — Nvidia being the most prominent case — had reached multiples that implied extraordinary long-term growth. These valuations created positive feedback loops: high valuations enabled cheap equity issuance, which funded continued capital expenditure, which maintained the narrative of competitive necessity, which supported continued high valuations. The feedback loop could sustain itself as long as the underlying business metrics continued to improve — but required specific future outcomes to justify present-day prices.
Third, corporate capital expenditure plans by the major hyperscalers had accelerated to levels that required either substantial revenue growth from AI-enabled services or willingness to operate at lower returns on invested capital than historical norms. The announcement of capital plans was credible because the companies had strong balance sheets. But the assumption that AI-enabled services would grow fast enough to justify the investment within the timelines embedded in the capital plans was exactly the kind of forward-looking assumption that had proven optimistic in previous infrastructure cycles.
The Expansion of Power
That is why this article matters. It gives the AI cluster a stronger crisis vocabulary without falling into empty doomerism, and it helps readers separate a useful technology thesis from a potentially fragile capital-market structure.
The data center debt machine had already been building for years before the AI acceleration — the fundamental business of building and financing large-scale computing infrastructure had developed its own credit structures, lease-back arrangements, and private capital channels. The AI cycle accelerated the pace of investment into an existing financing infrastructure, which created concentration risk in the credit channels that had scaled to support data center construction.
The BIS warning was not that AI investment was wrong. It was that the speed and volume of capital deployment, combined with the compressed timelines for return expectations, created conditions where a disappointing quarter or a slower-than-expected commercial adoption curve could produce a significant correction across multiple asset classes simultaneously — equity valuations, private credit portfolios, and AI-adjacent corporate bonds could all face repricing at the same time if the growth trajectory proved slower than the financing assumptions required.
The scale of the AI infrastructure investment race meant that the financing structure had become systemic in a way that earlier technology infrastructure cycles were not — because the investors, lenders, and equity market participants were now spread across pension funds, insurance companies, sovereign wealth funds, and retail investors through public equity exposure. A correction would not be contained within specialist technology investors.
The Hidden Strategy Behind the Fortune
The hidden strategy behind the fortune was showing that the danger in an infrastructure boom often sits in the financing assumptions around future returns, not in the underlying technology itself.
The pattern identified by the BIS was not unprecedented. The dot-com boom of 1999-2000 involved real technologies — the internet genuinely was transforming commerce, communication, and information access. But the financing assumptions embedded in valuations implied revenue growth rates that required the transformation to happen faster and more completely than the technology’s adoption curve could support. When the growth rates proved achievable but slower than assumed, the valuation correction was severe even though the underlying technology continued to develop.
Telecom infrastructure investment in the late 1990s showed a similar pattern. The build-out of fiber optic networks was appropriate given the long-run demand for bandwidth — bandwidth demand did eventually grow to fill the capacity that was installed. But the speed of construction created overcapacity relative to near-term demand, which produced severe price deflation in bandwidth markets and destroyed the financial structures that had financed the build-out, even though the infrastructure itself was ultimately useful.
The lasting lesson is about how expected future cash flows, debt tolerance, valuation support, and infrastructure financing discipline became levers that could amplify both the gains from a genuine technology opportunity and the losses from a financing structure that had optimistically priced that opportunity.
The Cost, Risk, or Collapse
If expected returns disappoint, the damage can spread through lenders, suppliers, private-credit vehicles, and valuation logic long before the underlying technology disappears.
The BIS warning was prospective rather than diagnostic — a description of conditions that could produce systemic stress, not a prediction of imminent collapse. As of 2024, the credit quality of the major AI infrastructure investors remained high, the revenue growth of the hyperscalers continued to support investment plans, and the commercial applications of AI were generating genuine enterprise revenue. The warning was about what would happen if these conditions changed.
The scenarios that could trigger the stress the BIS described included: a significant slowdown in enterprise AI adoption that would stretch the timelines for hyperscaler revenue recovery on infrastructure investment; a technological development that reduced the computational requirements for large language model inference, which would reduce the demand for the hardware that current infrastructure was built to supply; or a credit market repricing that raised the cost of the private-credit instruments financing smaller AI infrastructure projects.
None of these scenarios required AI to fail as a technology. They only required AI to develop on a slower or different path than the financing assumptions embedded in the current capital cycle. That distinction — between technology risk and financing risk — was the core of the BIS warning.
Lessons for Modern Business Readers
1. Technology plausibility makes financing risk harder to see
The strongest infrastructure bubbles combine genuine technological promise with financing assumptions that require the technology to develop on an optimistic timeline. The plausibility of the technology narrative makes it easier for investors to accept compressed return timelines and high leverage ratios. This is the specific mechanism the BIS identified: not that AI is unlikely to be transformative, but that the financing assumptions around the transformation timeline may be optimistic.
2. Private credit opacity creates systemic risk
Private credit markets expanded significantly into AI infrastructure financing because they could provide capital to companies and projects that did not meet public debt market standards. But private credit typically involves less transparency, less liquidity, and harder-to-assess valuations than public markets. When private credit portfolios contain similar exposures to the same cycle, repricing can be simultaneous and correlated rather than gradual and distributed.
3. Competitive necessity is not the same as financial viability
Companies invested in AI infrastructure partly because the competitive dynamics of the AI market made early investment appear necessary — the risk of falling behind competitors was real and proximate, while the risk of over-investing was future and probabilistic. This asymmetry of perceived risk is a consistent feature of infrastructure boom cycles. Each individual investment decision is rational; the aggregate result can be excessive.
4. Separate the technology thesis from the capital cycle thesis
The BIS warning does not require skepticism about AI’s long-run economic impact. It requires separating the technology thesis — “AI will eventually transform significant parts of the economy” — from the capital cycle thesis — “the current financing structures will generate returns within the timelines they assume.” These are independent claims that can have different outcomes.
5. Infrastructure value and infrastructure financing can diverge dramatically
The British railway network built during Railway Mania remained economically valuable for more than a century. Most of the investors who financed its construction lost money. The existence of long-run value in an infrastructure system does not protect near-term financing structures from the consequences of optimistic return assumptions. The infrastructure can be permanently useful while the financing structures that built it are temporarily impaired.
Conclusion
Seen clearly, this is not a story about whether AI is real or whether AI infrastructure will eventually be necessary. It is a story about the gap between technology value and financing assumptions — and about why the BIS chose to flag that gap specifically in the context of AI infrastructure investment.
That is why the article belongs inside the Hidden Fortunes ecosystem. It connects the Railway Mania crisis framework and the data center debt machine analysis to current institutional concern about the financing structure of the AI infrastructure cycle — giving readers a framework for tracking the difference between technological development and capital market assumption.
Book Recommendation
For readers who want the strongest next step, start with This Time Is Different by Carmen Reinhart and Kenneth Rogoff. It is the most rigorous historical documentation of how financial crises recur across eight centuries — showing that the belief that current conditions are uniquely immune to the patterns of previous cycles is itself the mechanism through which those patterns repeat.
