The artificial-intelligence infrastructure race might be entering an opaque phase. What began as a cash-flow-funded spending spree by the world's richest technology companies is increasingly becoming a debt-driven capital-expenditure cycle. A vast amount of capital is routed through leases, joint ventures, purchase commitments and private-market financing.
The tension is that the hyperscalers still look extraordinarily healthy. Companies such as Alphabet Inc.
But, according to Robin Wigglesworth, author and editor of the FT Alphaville - official balance sheets are no longer telling the full story.
"It's one of the biggest capital markets events of our lifetimes," Wigglesworth said in a recent interview.
For years, the hyperscalers could fund data centers from their core businesses because "Google search, Amazon, Facebook itself just prints money." Now, he said, "the scale is just becoming so massive that they've increasingly turned to the debt markets."
The Hidden AI Balance Sheet
The clearest example is Meta's Hyperion data center in Louisiana. Rather than simply issue debt itself, Meta formed a joint venture with Blue Owl called Beignet, took a 20% stake and guaranteed that it would lease the facility for at least 20 years. That structure allowed Beignet to issue $27 billion of bonds while keeping the debt off Meta's balance sheet.
"It doesn't come up as a bond or a debt or a loan for Meta," Wigglesworth said. "But, of course, it's on the hook for paying this lease for 20 years."
Goldman Sachs has estimated that hyperscalers carry about $1.5 trillion in lease commitments. Roughly $1 trillion relates to leases that have not yet commenced and therefore are not recognized as conventional liabilities under U.S. accounting rules. They reside largely in disclosure footnotes until the lease begins.
The companies are also accumulating contractual commitments for chips, computing capacity, cooling systems and energy. Those purchase obligations have reached nearly $1.5 trillion, led by Alphabet, which disclosed more than $800 billion in future commitments.
"These are financial liabilities that are in many cases extremely hard to squirrel out of," Wigglesworth said. "They kind of walk, talk and quack a bit like debt, but they don't actually appear as debt."
Compute Becomes Collateral
The next stage is the financialization of compute itself. NVIDIA Corporation
The comparison to commercial real estate or energy infrastructure is tempting. But compute has a shorter and less certain useful life than a pipeline, power plant or office building. Chips depreciate, fail, require replacement and can be superseded by new generations of hardware.
"Just because you say something is an asset class doesn't make it so," he said. Chips degrade, require maintenance and can become obsolete quickly. Those risks can be priced, he added, but the sheer volume of money and the prevailing fear of missing out make the cycle more fragile.
William Lee of Global Economic Advisors is more direct. "I want to be the person that raises that awful word leverage," he said on CNBC. The parallels, he argued, are with captive finance models such as GMAC in autos - except now the financing is tied to compute.
"We're seeing NVIDIA providing financing for compute," Lee said. Private credit funds are taking on more exposure, while banks are indirectly connected through credit lines that provide liquidity to the system. If rates rise or liquidity dries up, he warned, "we start to have a crunch."
The risk is more concentrated among companies without the core cash-generating businesses of the biggest hyperscalers. Oracle Corporation is more indebted relative to revenue than its larger peers, while CoreWeave Inc. represents the more leveraged neocloud model: massive borrowings underwriting an AI buildout without the same cushion from legacy businesses.
"I'm not worried about Facebook and Alphabet or Amazon going bust," Wigglesworth said. "There will be, of course, in any cycle a few extreme outliers that just borrowed way too much money."
Why the Boom May Last
Marko Papic of BCA Research expects AI infrastructure to eventually face overbuilding, as has happened with previous investment booms in areas such as canals and fiber-optic networks.
However, Papic does not see a bust as necessarily imminent. Falling token costs and the rise of open-source models are making AI cheaper to deploy. That could broaden adoption, increase demand for computing capacity and support further data-center investment.
Papic argues that lower AI costs are helping extend the capital spending cycle because data centers can still generate attractive returns and support viable business cases. The key warning sign would emerge when the pace of AI capital spending growth begins to slow and earnings stop delivering positive surprises.
