The AI Bubble Might Burst in Debt Markets, Not Stock Prices

The BIS question is not "Is Nvidia's P/E too high?"

The AI bubble debate usually starts with stock prices. How many multiples has Nvidia gained? Are Big Tech valuations above historical averages? How large would a correction be?

But the financial stability question that Pablo Hernandez de Cos, General Manager of the Bank for International Settlements, raised in a speech on September 10 is different. Money flowing into AI infrastructure is growing fast, faster than companies can fund it from their own cash. The question is what kind of linkages that creates inside the financial system.

If a stock falls 30 percent, the damage can end with someone's paper loss. But if a data center worth 10 billion dollars was built with borrowed money, and rental income and compute revenue come in below expectations, the fallout can spread to lenders, private credit funds, construction firms, server makers and power suppliers. A stock bubble and a credit bubble get lumped under the same word, but they behave differently.

Split the money into two kinds

The first kind comes from outside the AI industry. A company buys Copilot seats. A developer pays for Claude or OpenAI API usage. A consumer pays for an AI subscription. An advertiser spends more on AI powered search or recommendation services. This is external demand, paid by end customers who feel they are getting value.

The second kind moves inside the AI ecosystem itself. A chipmaker or a hyperscaler takes an equity stake in an AI lab, and the lab uses that money to sign contracts for GPUs and cloud capacity. A BIS working paper that mapped disclosed transactions found roughly 46 billion dollars in equity investments and 879 billion dollars in multi-year purchase commitments woven through this network.

None of this structure is automatically bad. It is common in early-stage industries for suppliers to help finance their customers' growth. Automakers help fund dealer financing. Telecom equipment makers help carriers pay for the gear they buy.

But it adds one more question investors need to ask. Not just "did the GPU sell?" but "how much of the money coming in for services built on that GPU is actually from an external customer?" If internal financing and purchase commitments grow much faster than final demand, the industry's growth rate can look strong while the economic quality of that growth is weak.

Circular financing does not mean fake revenue

The easiest misunderstanding to fall into here is this. If Nvidia or a hyperscaler invests in an AI company, and that company then buys GPUs or compute with the money, the resulting revenue is not automatically fake in an accounting sense. Real cash can change hands, and real equipment and services can be delivered.

The issue is independence. As the share of financing that a supplier arranges for its own customer, and that then flows back as the supplier's revenue, grows larger, it becomes necessary to separate demand that end customers pay for with money they earned themselves from demand created because the ecosystem supplied the funding first.

So circular financing is less an accusation of accounting fraud and more a lens for judging the quality of demand and the depth of financial interconnection.

Why debt changes the nature of the risk

If a data center is built with 100 units of a company's own money and returns disappoint, the loss mostly lands on shareholders. But if the mix is 30 units of equity and 70 units of debt, the story changes. Even if revenue comes in below expectations, interest still has to be paid on a fixed schedule.

AI infrastructure requires especially large upfront investment. Money goes first into GPUs, networking, cooling, power connections, land and buildings, and revenue is recovered over many years afterward. That means cash flow can come under stress if expected utilization rates fall or compute prices drop quickly.

An even harder problem is collateral. Unlike a generic office building, assets built around a specific power architecture, cooling system and GPU generation may be difficult to sell at full value in a fire sale. The BIS working paper models how this kind of "specialized asset fire sale," combined with debt, can amplify losses.

The "50 percent overinvestment" figure is not an audit of actual capex

The most attention-grabbing number in the BIS working paper is roughly 50 percent. The paper calculates that in a near winner-take-all market, companies racing to get ahead of rivals can invest about 1.5 times the socially efficient level. Under conditions where demand responds less to price changes, that figure can grow to roughly three times.

This number should not be read as "half of current AI capex is being wasted." It is not a survey that counted unnecessary equipment data center by data center. It is the output of an economic model built on assumptions about competition and financing structure. The working paper itself states explicitly that it reflects the authors' research and not the official view of BIS member central banks.

What the model is really pointing to is incentives, not the number itself. If a company believes the market leader will capture most of the profit, each competitor has an incentive to choose "build too much" over "build a little less." What is rational for an individual company can add up to industry-wide oversupply.

Still, this is not automatically the dot-com bust or the 2008 crisis

The counterargument is strong too. The center of current AI investment includes companies with enormous cash flow, Microsoft, Alphabet, Amazon and Meta among them. AI demand is not built purely on expectation either. Revenue and productivity effects are already showing up in enterprise software, cloud, advertising, coding and search.

Data centers also are not one-shot assets like railroads laid down and finished. Even if models change, much of the power, cooling and network infrastructure can be reused for other computing demand. And a rise in private credit does not automatically mean a repeat of the 2008-style financial crisis.

So reducing the BIS argument to "AI is a bubble" misses the more important point. The sharper question is this: even if AI succeeds as a technology, will every data center and every GPU being built now actually generate the returns investors expect?

Insight Times Editorial Desk