Global technology companies jumping into the artificial intelligence (AI) race are expected to pour $7 trillion (about 9,700 trillion won) into data centers by 2030, deepening Wall Street's concerns. Even cash-rich big tech companies are finding it hard to shoulder the massive investment costs with their own funds alone, leading to increased reliance on external financing such as bonds and loans. If the value of data centers and AI Semiconductor chips falls quickly or AI demand falls short of expectations, the investment risk could spread from tech companies to the financial sector.
On the 26th (local time), the Financial Times (FT) reported that technology companies are expected to commit about $7 trillion (about 9,700 trillion won) to data centers by 2030. In particular, even big tech companies holding tens of billions of dollars in cash are finding it difficult to build the necessary infrastructure with internal funds alone, prompting a flood of financing from the financial sector.
AI infrastructure also appears to be taking up a larger share of the U.S. economy. Stijn Van Nieuwerburgh, an economist at Columbia Business School, estimated that future AI infrastructure buildout will account for about 2.8% of U.S. economic output. The so-called four "hyperscalers" — Amazon, Microsoft (MS), Alphabet, and Meta — have lease agreements signed since the AI boom that, including contracts that will take effect in the future, exceed $1.5 trillion (about 2,079 trillion won).
The problem is that it is hard to guarantee that the data centers built with massive borrowed funds and the AI Semiconductor chips inside them will maintain their current value throughout the financing period. With Nvidia rolling out new chips that improve performance every year, existing chips could lose value quickly within a few years. For financial firms that lent against the chips installed in data centers as collateral, that means bearing the risk that the collateral value will fall before the loans are repaid.
In particular, the long-term value of data centers themselves is uncertain. Current data center investment assumes that massive AI compute demand will continue, allowing the facilities to be used continuously. But technological advances could reduce the computing resources needed to run AI models, or supply could exceed demand, making it impossible to rule out the possibility that data center values will fall faster than expected.
This uncertainty has begun to be reflected in actual financing expense. In April, U.S. data center firm QTS issued $4.6 billion (about 6.3 trillion won) in bonds to build a data center in Fayetteville, Georgia. Orders at the time were about three times the issuance.
But investors have since begun to focus on the fact that the bond principal will not be fully repaid before the initial lease at the data center expires. If a new tenant cannot be found after the lease ends, bond repayment could be jeopardized. In fact, the yield on bonds newly issued by QTS this month exceeded 7.2%, a sharp rise from 5.7% in April. FT said that is similar to the financing expense for companies with junk ratings (high-risk credit ratings below investment grade).
Major banks are already looking for ways to reduce the risks concentrated in data centers. JPMorgan Chase, Morgan Stanley, and Sumitomo Mitsui Banking Corporation (SMBC) considered "synthetic risk transfer (SRT)" transactions, which keep the loans on their books while passing some loss risk to external investors. In some data center projects, special purpose vehicles (SPVs) are being used to keep operations off big tech balance sheets, with external investors bearing the risk.
The risks the financial sector must weigh are not limited to asset value declines. As data centers consume massive amounts of power and water, community opposition is intensifying across the United States. In a Gallup poll in March, seven out of 10 Americans said they opposed data centers being built in their area. In Virginia, a data center project backed by Blackstone was canceled after facing local opposition and lawsuits, and some local governments have moved to temporarily halt new data center development. If permits are revoked mid-project, financial firms that have already lent money could be left holding the loss risk.
It is not easy to cover all the risks with insurance. As cases of a single data center costing more than $10 billion (about 14 trillion won) to build have become common, the amount of insurance needed for natural disasters or power outages has risen sharply. According to FT, for Meta's $14 billion (about 19 trillion won) data center under construction in Texas, only part of the total is covered by insurance, leaving the company exposed to losses of several billion dollars if a disaster or outage occurs.
The biggest variable here is AI demand. Current data center investment assumes that AI will significantly boost productivity and corporate profits and that the resulting massive compute demand will persist for a long time. But if AI models become more efficient faster than expected or if data center supply exceeds demand, the facilities that swallowed massive amounts of capital could remain as "stranded assets," failing to generate sufficient revenue.
Andy DeVries, a utilities analyst at CreditSights, projected that AI compute supply capacity will begin to outpace demand after 2029. If the frenzy for data center investment cools, the impact likely will not be limited to big tech and the financial sector. Local governments that offered tax incentives to attract data centers could face revenue shortfalls, and the data center-related expenses left with local communities could translate into higher utility bills for ordinary consumers, FT said.