The AI Hegemony Race Turns into a "Race of Debt"—BIS Warns of the Next Tech Bubble

The competition in generative AI has shifted from software development focused on performance to a massive capital investment race, vying for power, semiconductors, data centers, communication networks, and funding.

The scale is no longer just a growth story of a single industry. The Bank for International Settlements (BIS) analyzes that the five major U.S. hyperscalers are expected to invest over $1 trillion in AI-related infrastructure from 2025 to 2026. The investment amount surpasses the growth in profits and free cash flow, with some companies increasingly relying on external funding such as corporate bonds and loans. (Sources 2, 3)

BIS is not warning that "AI is useless." On the contrary, the problem is the opposite. Because AI is likely to become a core technology of the future, companies cannot stop investing. Companies that fall behind may be disadvantaged in model performance, users, data, and computing power, potentially losing long-term market dominance.

While investments may be rational for individual companies, they become excessive for the industry as a whole. This contradiction is at the core of the current AI boom.

"Winner-Takes-Most Market" Leads to Overinvestment

In the AI market, companies with the largest computing resources and the most users are considered to have an advantage. This is because they can operate large models, improve with the data obtained, and attract more users in a cycle.

From a managerial standpoint, it seems more rational to secure computing power, even if it means stretching finances, rather than protecting the financials by limiting investments. If a company becomes cautious while competitors continue to invest, it risks falling out of the future winner's circle.

However, if all companies operate under the same logic, multiple data centers will be built in the same region, and computing power aimed at the same customers will overlap. If only a few companies ultimately achieve high profits, some of the funds invested by the remaining companies may not be recovered.

BIS's research model indicates that AI investment could exceed socially efficient levels by about 50% even under conservative assumptions. In conditions where demand is less responsive to price changes, overinvestment could reach about three times the efficient level. The larger the investment scale, the deeper the downturn tends to be when it reverses. (Source 2)

The important point here is that even if AI usage increases, it doesn't necessarily mean that investors' profits will increase. If competition drives down the price of AI services, while companies and consumers benefit, the profit margins of those who built the data centers will decline.

A situation could arise where it is a success socially but a failure as an investment project.

From a Cash Competition to a Debt Competition

For years, major IT companies have leveraged the abundant cash generated from their core businesses as a strength. However, as AI investments surge and large-scale projects span multiple years simultaneously, internal funds alone become insufficient.

According to BIS, the amount of corporate bonds issued by major hyperscalers exceeded $100 billion in 2025. Additionally, structures where data centers are owned by special purpose companies or joint ventures, with technology companies entering into long-term leases or computing power purchase agreements, are expanding. (Source 4)

In this structure, much of the related debt does not appear directly on the balance sheets of the technology companies themselves. BIS describes this state, where obligations economically akin to debt are placed on external entities, as "shadow borrowing."

Even if the accounting location changes, the burden does not disappear when demand vanishes. This is because companies with long-term contracts, data center owners, private credit funds that provided funding, banks, and insurance companies are intricately connected.

The balance of private credit loans to AI-related fields has increased from almost zero to over $200 billion, with its share of the total rising to nearly 8%. BIS suggests it could swell to $300 billion to $600 billion by 2030. (Source 5)

Currently, the AI-related ratio of an average fund is limited, but the problem is not visible in the average alone. If risks are concentrated in specific lending companies, construction companies, power operators, or semiconductor companies, financial difficulties in one company could ripple into another's cash flow.

Even if AI is Real, a Bubble Can Occur

The claim that "AI is a revolution comparable to the internet, so it's not an investment bubble" has a pitfall.

Historically, large-scale bubbles did not occur solely with worthless technologies. The canals and railroads of the 19th century, the power grids of the early 20th century, and the internet of the 1990s all significantly changed society. At the same time, they led to overconstruction and capital inflow that anticipated future demand, causing losses to numerous companies and investors. (Source 3)

Even after the dot-com bubble burst, the internet did not disappear. Rather, the next generation of companies utilized the now cheaper communication networks and server infrastructure to cultivate new services.

The same could happen with AI. While the technology spreads and enhances societal productivity, investors who bought stocks at current high prices or lenders who financed unprofitable data centers might incur losses.

In other words, the "AI revolution" and the "AI investment bubble" are not mutually exclusive concepts. Both could be true simultaneously.

On Social Media, "Crisis," "Overreaction," and "Infrastructure Remains" Intersect

Reactions on public social media are largely divided into three directions. However, the following is a qualitative organization of public posts and does not represent overall public opinion.

The first is a reaction concerned about debt and the opacity of information.

Comments on BIS's LinkedIn post pointed out that if current servers and semiconductors become obsolete before the debt is repaid, both collateral value and future earnings could weaken. There was also a view that the stage where technology investments shift from being covered by cash to borrowing is when macroeconomic risks quietly accumulate.

The common stance is that one should look at "who is lending, under what conditions, and how much," rather than the performance of the models. (Source 7)

The second is a rebuttal to excessive pessimism.

Unlike the deficit companies of the dot-com era, current investment entities are reaping massive profits from cloud, advertising, e-commerce, and business software. Even if there are some investment failures, it does not necessarily mean they will immediately fall into insolvency.

In a survey of institutional investors introduced by Reuters, while 82% viewed AI as the most crowded trade, about half said it was not a bubble. The caution and bullish stance among market participants are both intensifying simultaneously. (Source 6)

The third is a long-term optimistic view that "even if it collapses, the infrastructure remains."

In discussions on Reddit, similar to the investment booms of railroads, electricity, and the internet, there was an opinion that even if investors temporarily incur losses, the constructed infrastructure will be utilized for the next economic growth. If surplus computing power is released cheaply, AI usage could spread to small and medium-sized enterprises, universities, and research institutions. (Source 8)

However, not all data centers can be easily repurposed. If power contracts, location, cooling methods, semiconductor generations, and communication connections do not fit the purpose, the building may remain but not generate high revenue.

The fact that infrastructure remains in society and that investors can recover their principal are separate issues.

The Trigger for Collapse May Not Be "AI Failure"

For the investment boom to reverse, AI does not need to suddenly become unusable.

A slight drop in the expected sales growth rate is sufficient. If demand shifts from expensive cutting-edge models to cheaper small or open models, the unit price of computing power and profit margins will decrease. Even if companies continue to use AI, the assumptions for investment recovery could collapse.

Shortages of power, transformers, construction personnel, and advanced semiconductors are also double-edged swords. While supply constraints drive up costs, they also prompt companies to secure future capacity early. If demand forecasts are revised downward after securing power or computing capacity through long-term contracts, only fixed costs remain. (Source 3)

Rising interest rates and deteriorating credit markets are also important. Funds raised through equity have no repayment deadline, but corporate bonds and loans have interest payments and maturities. Even if monetization is delayed by a few years, if refinancing conditions worsen, companies may be forced to reduce investments.

Furthermore, if one company reduces capital investment, the impact will ripple through to semiconductor manufacturers, construction companies, power generation and transmission companies, and data center owners. BIS points out that not only AI companies but also design, procurement, and construction contractors with relatively weak financial bases are susceptible to the effects of investment reversal. (Source 3)

The Financial Industry Supports the Boom While Exercising Caution

The financial industry, recognizing the dangers of AI investment, is simultaneously a beneficiary of the boom.

Data center financing, corporate bond issuance, equity fundraising, mergers and acquisitions, and the formation of special purpose companies bring significant fee income to banks and investment firms. Major financial institutions view AI-related capital investment as a multi-year "super cycle." (Source 9)

If companies stop investing, they lose the competition; if banks hold back on lending, they lose deals to others; if investors don't hold AI stocks, they underperform the market average.

Even if each party believes they are acting rationally, if everyone fears the "risk of not participating," investments and loans concentrate in the same direction.

Bubbles do not necessarily occur because no one knows the danger. They occur when everyone knows the danger but cannot exit first.

Five Numbers Investors and Companies Should Watch

When assessing the future AI market, looking only at model performance and user numbers is insufficient.

First, the growth of AI-related sales relative to capital investment amounts. Even if sales are increasing, if investment amounts are growing faster, the payback period will be longer.

Second, free cash flow. Even if operating profit is positive, if cash continues to decrease after deducting capital investments, additional borrowing or capital raising will be necessary.

Third, the credit spread of corporate bonds and refinancing terms. Even if stock prices are strong, if the interest rates demanded by the bond market are rising, lenders are reassessing the risk.

Fourth, fixed payment obligations, including off-balance-sheet contracts. Without summing up leases, minimum purchase quantities, power purchase agreements, capacity reservations, and guarantees, the actual debt is not visible.

Fifth, the concentration of customers and suppliers. Semiconductor, power, and construction-related companies dependent on the investment plans of a few companies are more likely to see a slowdown in orders directly affecting sales.

The Question is Not Technology, but Time

The potential for AI to change the world is high. However, the speed of technology diffusion and the repayment speed required by financial contracts do not necessarily align.

Even if an infrastructure generates huge profits in 10 years, if refinancing fails in 3 years, ownership changes. Even if a technology leaves value in society, the initial investors may not profit.

BIS's warning is not a call to stop AI development. It points out that the competition to become future winners leads companies to excessive upfront investment, and the debt and mutual investments supporting this create new contagion paths into the financial system.

What truly needs to be observed in the current AI market is not the simple question of "Is AI real?"

Who is investing, who is borrowing, who is guaranteeing, and who will bear the losses if the expected profits do not materialize? Without understanding this structure, the risks swelling behind the shine of technology remain unseen.

Will the AI boom transition into sustainable growth, or will it become the entrance to a long investment recession?

The turning point is not determined solely by the winners of the performance race. It is the difference between the time debt can wait and the time until profits are generated.

Sources and References

Source 1: Financial Post/Bloomberg News

Reports on the potential for competition among companies over AI infrastructure to develop into a scale surpassing past technology investment booms, affecting the entire market through borrowing and inter-company financial relationships.

https://financialpost.com/investing/ai-investment-race-could-turn-debt-fuelled-boom-to-bust

Source 2: BIS Working Paper "The AI investment race"

Analyzed competition models where winners capture most of the market, investments exceeding efficiency levels, and financial vulnerabilities due to debt and cyclical investment relationships. Used estimates showing overinvestment of about 1.5 times under conservative conditions and about 3 times under certain conditions. Note that the paper reflects the authors' views and does not necessarily represent the official views of BIS or its member central banks.

https://www.bis.org/publ/work1367.htm

Source 3: BIS "Annual Economic Report 2026 — Progress and peril"

Used for analysis of over $1 trillion AI-related infrastructure investment by the five major hyperscalers from 2025 to 2026, supply constraints, financial and real economic impacts of investment reversal, and comparisons with past investments in railroads, electricity, and dot-coms.

https://www.bis.org/publ/arpdf/ar2026e1.htm

Source 4: BIS "Financing the AI infrastructure boom: on- and off-balance sheet borrowing"

Used for analysis of over $100 billion corporate bond issuance by hyperscalers in 2025, data center investments using special purpose companies and joint ventures, long-term leases, and the structure of off-balance-sheet "shadow borrowing."

https://www.bis.org/publ/qtrpdf/r_qt2603u.htm

Source 5: BIS Bulletin "Financing the AI boom: from cash flows to debt"

Used for analysis of private credit loans to AI-related fields exceeding $200 billion, expansion to nearly 8% of total loans, forecasts of loan balances by 2030, and analysis of AI-related loan scale, duration, and interest rate conditions.##