Sponsored

The AI buildout has a balance-sheet problem, and the fix is changing who owns the asset.

For three years, hyperscalers paid for data centers the way they pay for everything: out of operating cash flow, on the corporate balance sheet, funded by businesses that throw off tens of billions in free cash. That worked while the bill was a rounding error against revenue. It is not a rounding error anymore. Goldman Sachs’s baseline model puts AI capital expenditure at roughly $765 billion in 2026, scaling toward $1.6 trillion a year by 2031 — about $7.6 trillion cumulatively across the period. Planned hyperscaler spending on AI and data centers tops $5 trillion by 2030 on the same estimates. No four balance sheets, however strong, absorb that without hitting concentration limits, credit-rating pressure, and the simple constraint that a single asset class should not eat the entire capex budget.

So the financing stack is rotating. The capital is increasingly coming from outside the hyperscaler — from private credit, infrastructure funds, and real-estate capital that treat a gigawatt data center the way they treat a power plant or a fiber network: a long-duration, contracted asset financed mostly with debt. Goldman’s research team frames the shift bluntly: private-market infrastructure and real-estate funds will play a growing role in data-center financing, and while AI capex has been internally funded “to date,” the rising reliance on debt is the thing to watch in 2026. As of May 2026 there were 695 infrastructure funds in the fundraising market chasing an aggregate $555 billion — a war chest aimed squarely at this asset class.

The deal template

The clearest signal is in the deals themselves. In October 2025, Meta formed a $27 billion joint venture with funds managed by Blue Owl Capital to build its Hyperion campus in Richland Parish, Louisiana. The structure is the point: Blue Owl owns 80 percent and Meta 20 percent of a special-purpose vehicle, arranged by Morgan Stanley, carrying $27 billion of debt against $2.5 billion of equity. PIMCO anchored the lending. The debt matures in 2049, is fully amortizing, and carries an A+ rating from S&P. Meta leases the facilities back on an initial four-year term and took a roughly $3 billion one-time distribution out of the JV at close.

Read that again. Meta gets the compute, keeps the leverage off its own balance sheet, and gets paid to do it. The asset sits in an SPV that private capital owns and lenders finance against contracted lease cash flows. It is the project-finance playbook that built toll roads and merchant power, repointed at GPUs.

Hyperion is not an outlier; it is becoming the template. Recent private-capital-led AI data-center projects cluster in the tens of billions:

Private credit alone deployed more than $15 billion into AI data centers, GPU clusters, and compute infrastructure in February 2026, a single month. The Aligned Data Centers acquisition — a roughly $40 billion deal pulling in Nvidia, Microsoft, BlackRock, and xAI — and the $40 billion-plus Stargate financings show the same pattern at the top of the market: equity sponsors assemble the project, then term it out with institutional debt rather than fund it from cash.

Self-build versus rent-the-balance-sheet

The economics explain the rotation. A hyperscaler self-build is the cheapest nominal cost of capital — these are some of the most cash-rich companies on earth — but it is the most expensive in scarce balance-sheet capacity. Every dollar of data-center capex is a dollar not available for buybacks, a notch of concentration risk, and a line item that rating agencies and equity investors increasingly mark against. The private route inverts that. The lease premium and the lenders’ coupon cost more in cash terms, but they buy off-balance-sheet treatment, faster construction, and risk transfer to counterparties who actively want twenty-five-year amortizing infrastructure paper.

That paper is the prize. An A+ rated, fully amortizing data-center loan maturing in 2049, backed by a lease from one of the world’s strongest credits, is exactly the asset insurers and pension funds need to match long-dated liabilities. The Bank for International Settlements flagged the flip side in March 2026: a growing share of these obligations are “economically akin to debt but largely reside outside corporate balance sheets” — shadow borrowing that gross bond issuance, itself topping $100 billion in 2025, only partly captures. The risk does not disappear when it leaves Meta’s balance sheet. It migrates to private credit funds, their insurance-company backers, and ultimately the savers behind them.

What it means

For the hyperscalers, this is a feature, not a retreat. Off-loading the asset preserves credit ratings and capital flexibility while keeping the compute. Expect the JV-and-leaseback structure to spread from greenfield campuses to power shells and even existing facilities.

For investors, the AI trade is quietly becoming a credit trade. The marginal dollar funding the buildout is increasingly a private-credit dollar, which means the cycle’s stress will show up first in fund marks, lease-coverage ratios, and the assumptions baked into 2049-dated amortization schedules — not only in Nvidia’s order book. The bull case for private credit is that it has found a vast, investment-grade, long-duration asset class at the exact moment it raised half a trillion dollars to deploy. The bear case is concentration: the same handful of tenants underwrite most of the leases, and the entire structure assumes AI demand holds long enough to amortize debt that comes due more than two decades out.

The data-center groundbreaking photo is the wrong thing to watch. The capital structure underneath it is where the AI buildout will be won or lost.

Finance & Markets Correspondent
Covers: Finance, capital markets, technology investing

David Whitmore covers the intersection of capital and code — the funding rounds, market structures and policy moves that shape how money flows through the technology economy.