Everyone talks about how much AI costs. Fewer people ask who's lending the money, and where that debt sits. A paper by Columbia Business School's Stijn Van Nieuwerburgh, presented at Brookings in September 2026, maps it out. This piece is about the financing structures. For whether the spending pays back, see the AI bubble capex math.
The size of it
From Brookings' Oct 8, 2026 summary:
- U.S. AI investment in data centers, power and chips could total $10.3 trillion from 2025 to 2032.
- That's 3.63% of GDP per year on average.
- For comparison, he puts the railroad boom at about 2.2% of GDP at its peak (about 1870 to 1890), highways at about 1.1% in the 1950s, and the telecom and fiber boom at about 1.1%.
On the Brookings podcast, host Aaron Klein noted that companies like Meta and Microsoft are now spending more on AI than they generate in operating cash flow. That gap is what the financing fills.
Where the money comes from
The paper says a growing share of capital comes from outside the big tech companies: real estate developers, infrastructure funds, private credit and structured finance vehicles. "In many cases, the user of compute is not the owner of the underlying assets."
The main tools:
- Leases. The tech company rents the building instead of owning it.
- Joint ventures. A partner buys most of the project's equity.
- Private credit. Loans from funds, not banks or public bond markets.
- Special-purpose vehicles (SPVs). A separate company whose only asset is the data center (or the chips) and which borrows against it. More in our SPV explainer.
The template: Meta's Hyperion
The paper uses Meta's Hyperion campus in Richland Parish, Louisiana:
- Meta sold an 80% equity stake to private credit firm Blue Owl for about $2.5B.
- The joint venture, Beignet, raised $27B of debt in October 2025, rated A+ by S&P, one notch below Meta.
- That's about 90% debt ($27B against roughly $30B of asset value), far above a typical investment-grade company.
On the podcast, Van Nieuwerburgh explained the leases: Meta signed five consecutive 4-year leases, and it can walk away every four years, but then owes a residual value guarantee (a payment) to the structure. Only the current lease shows up on Meta's balance sheet. The future leases don't, until they start.
Why that matters
His point isn't that the deal is bad. He said that as long as Meta doesn't go bankrupt, Hyperion should generate enough cash to repay the debt. The concern is visibility:
- Leverage sits in the SPV, so the tech company "looks like it's relatively under-leveraged."
- Lease and consolidation decisions are "judgment-heavy."
- If a tenant leaves and nobody else wants a GPU facility, the loss falls on the landlord and lenders.
- Hyperscalers still bear a lot of risk because they usually own the IT gear inside.
He draws a cautious parallel to pre-2008 off-balance-sheet bank vehicles, which ended up back on balance sheets. Brookings' summary: "premature to call this systemic risk," but "better measurement and disclosure should come while the industry's capital structure is still evolving."
What to watch
When Big Tech reports next week (dates and lines), look past capex. Check lease commitments, guarantees and joint-venture disclosures in the notes. NVIDIA's guarantee of up to $105B for a campus leased to OpenAI (NVIDIA preview) is the same theme.
