Meta’s US$27 billion (S$34.52 billion) Hyperion data centre illustrates a growing financial risk in the artificial intelligence boom: debt that sits outside a company’s balance sheet.
The facility is 20 per cent owned by Meta and 80 per cent by private credit giant Blue Owl and funded through bonds with maturities of about 24 years, sparing Meta most of the upfront cost.
Meta’s rent will help repay the bonds, but it has committed to only four years of rent. If Meta leaves during the first 16 years of operations and the property is worth less than an agreed minimum, it must make up the shortfall.
Meta could therefore face a substantial bill if the project falters, yet the tech giant has recorded no liability for the guarantee because it judges a payout is not probable. Still, its filings put its maximum exposure to the venture at a hefty US$46 billion, through investment, funding commitments, leases and guarantees.
This is just the tip of the Big Tech debt iceberg. Five US tech giants have about US$1.35 trillion in debt on balance sheets and a further US$1.65 trillion in commitments, mostly in chips and data centres. Big Tech has also guaranteed up to US$300 billion of other companies’ AI borrowing.
What’s happening with Meta and Hyperion echoes past events, critics say. Before 2008, banks set up separate entities known as conduits to buy mortgage securities using short-term borrowing. Banks promised to provide funding if those entities could not refinance. When investors stopped renewing their loans in 2007, banks had to step in and bring the assets and risks back onto their books.
Hyperion is sturdier because its long-term bonds are less vulnerable to sudden refinancing pressures. However, as with the conduits, a pledge means much of the risk still belongs to Meta.
Circular financing
A second trend adds to the concern: AI chip suppliers financing their own customers. Nvidia has invested US$30 billion in OpenAI, which buys its chips. It has also agreed to guarantee OpenAI’s 20-year data centre lease, up to US$105 billion for the first phase.
Critics call this circular financing. Nvidia’s backing helps OpenAI secure compute infrastructure and buy more Nvidia chips, which can make demand for chips look more self-sustaining. To be sure, Nvidia rejects the label, and supplier financing can indeed help a young industry take off before it generates enough revenues to fund itself.
But, like Hyperion, most of these deals rest on a comfortable assumption: that Big Tech is “too big to fall”, and strong enough to honour its promises whatever happens. Yet if OpenAI earns less than hoped, Nvidia could be hit three times: fewer chip sales, a less valuable stake in OpenAI and losses under its guarantees.
When the loan outlives the chip
AI lending also carries a problem of timing. It runs on three separate clocks: how long the debt runs, how long customer contracts last and how long the chips stay useful.
Consider CoreWeave, which rents out computing capacity powered by Nvidia chips. In August, it secured a roughly five-year loan backed by customer contracts averaging three years. To price in the risk of contracts going unrenewed, lenders are charging higher interest rates.
Here, there is one more complication: the chips themselves serve as collateral. If CoreWeave defaults, lenders can seize and sell them. But because that collateral loses value as more powerful, energy-efficient chips are produced, this additional risk must also be priced into the interest rate.
Correction or crisis?
All things considered, a sharp correction in AI investment is likelier than a financial crisis. Morgan Stanley expects Big Tech’s own cash to cover about half of the US$2.9 trillion data centre spend over four years. Most remaining debt is long-term, unlike the short-term funds frozen in 2008.
History also offers a useful comparison in the telecommunications boom in the late 1990s and early 2000s. Telcos borrowed heavily to build networks before demand caught up. Many subsequently went bankrupt, while equipment suppliers such as Lucent and Nortel suffered falling sales and credit losses after financing customers’ purchases. The sector endured a painful correction without bringing down the banking system.
A correction can become a crisis when losses spread through financial institutions that cannot see where the exposure lies. In 2008, uncertainty about mortgage losses made banks reluctant to lend to one another, turning losses in one market into a giant financial crisis.
That risk is not insignificant. Today, more AI financing sits with private credit funds, which lend outside the traditional banking system and generally disclose less than publicly listed companies. Banks also lend to these funds, while insurers have become significant investors in private credit. Regulators face the challenge of identifying who ultimately bears the risk and how losses could spread.
Many funds may have similar exposures to AI infrastructure and the same small group of major customers. Lenders could pull back simultaneously when deals sour. Nervous investors have already pushed several to limit withdrawals in 2026, including funds run by Blue Owl. Rising US interest rates add to the strain, especially for floating-rate loans such as CoreWeave’s. If losses spread to banks and insurers, credit could tighten for other businesses. A crisis could start outside the banks and still reach them.
Managing risks prudently
Singapore has vested interests in a continued AI boom. AI demand has boosted non-oil domestic exports over the past two years. DBS Bank, OCBC Bank and UOB finance regional data centres. Temasek has 6% of its portfolio in AI, aiming for up to 15% by 2031. A sharp correction could hurt exporters and jobs, reduce investment returns and strain bank borrowers.
Thus far, Singapore’s banking system appears resilient. The Monetary Authority of Singapore has assessed our banks to be capable of withstanding a severe AI downturn, though about a third of listed firms would come under financial strain in its corporate stress test.
While that is reassuring, containing the fallout also depends on whether companies, lenders and investors exposed to AI are asking the right questions: Who absorbs the loss during a default? Is incoming cash driven by the next funding round or by paying customers? Will the loan, the contract or chips run out first? And who else is exposed to the same customer?
Companies should report guarantees, triggers and maximum exposure together in plain language, and investors should treat these promises as real risks, not footnotes. Lenders must stress-test renewal assumptions against the rapid depreciation of technological collateral. Regulators must aggregate total exposures across banks, private funds and insurers, monitoring systemic linkages closely.
All this isn’t to say that a financial crisis is imminent. Such arrangements help share the huge upfront cost of this AI build-out. And, to be sure, in the Hyperion case, no rules were broken, the arrangement complies with accounting rules, and the disclosed figures are not evidence of impending loss. Moving long-term risk to long-term investors also aligns financial returns with risk appetite and debt tenure.
For this race to continue, we need watchful marshals to flag signs of danger, a safety car to slow the field, and medics ready with capital and cash, so that one crash does not end the race for everyone. But marshals cannot flag what they cannot see.
This commentary was first published in The Straits Times.