AI's Data-Center Debt Isn't Hidden. It Just Moved.

The short-form version of this story is that AI companies are running an Enron play: burying the debt they need to build data centers inside shell entities so it never lands where you'd look for it. It's a good hook. It's also wrong about the mechanics — and, more usefully, wrong about where the risk actually sits.

Start with the reframe. A special-purpose vehicle is not a lie. It's a location. It's a separate legal entity built to do one thing: own an asset, borrow against its cash flows, and wall the project off from its sponsors' other activities. Power plants, pipelines, toll roads, aircraft, and cell towers have been financed this way for forty years. The debt is real, and in today's deals it is disclosed. What's new isn't the structure. It's the scale, the speed, and the fact that the companies signing the contracts are increasingly financing one another.

So the honest question isn't “fraud or not fraud.” It's this: when you rebuild these companies' balance sheets to include what they've moved off them, and you discount the revenue that's really just their own capital coming back around, is the equity still worth what the market says? That's a valuation problem, not a morality play. Off balance sheet doesn't mean hidden. It means the balance sheet may no longer tell you where the risk ends.

Why AI companies use SPVs — and why a competent CFO would

The buildout is hard to overstate. A modern data center needs land, power generation and transmission, cooling, networking, buildings, and enormous quantities of specialized chips. It takes years and can cost tens of billions before it earns a dollar. Even the largest technology companies have good reasons not to fund every project straight off their own balance sheet.

The typical structure: a project company is formed; the tech company and outside investors put in equity; the project company borrows most of the construction cost, owns the facility, and leases it to — or sells its capacity to — the tech company, whose payments service the debt. The advantages are real and legitimate:

• It matches long-lived assets to long-term capital instead of funding a decades-long asset out of one year's cash flow.
• It brings in investors who specifically want infrastructure exposure — pension funds, insurers, private-credit funds — who are better suited to hold long-duration, asset-backed paper than a software company's shareholders.
• It ring-fences execution risk: cost overruns, permitting delays, and contractor claims stay inside the project.
• It preserves the parent's balance sheet and credit rating for the things where it actually has an edge — models, products, distribution — in a world where chip architecture changes faster than buildings do.

None of that is deceptive. It's recognizable project finance. The concern starts precisely when the outside capital appears to carry the risk legally while the tech company keeps carrying it economically.

Meta's Hyperion deal shows both sides at once

Meta's Hyperion campus in Louisiana is the cleanest illustration. In October 2025, Meta formed a joint venture with funds managed by Blue Owl Capital — Blue Owl holding 80 percent, Meta 20 percent — to develop the roughly $27 billion project, with the debt sitting in the venture rather than on Meta's books.⁶ Meta is the facility's tenant. Weeks later, Meta went out and raised roughly $30 billion in corporate bonds — capacity it plausibly still had because the data-center debt wasn't on its balance sheet.⁷

That's a legitimate allocation of ownership. It is not a clean transfer of economic risk — and the reason is in the fine print. Meta granted the venture a residual-value guarantee covering roughly the first 16 years: if Meta walks away at a lease non-renewal and the facility's resale value falls below a pre-agreed threshold, Meta covers the shortfall to the investors. The initial lease term is only about four years — strikingly short for a hyperscale asset, where ten years or more is standard.⁸

So the debt may not be Meta's legal debt, but Meta's credit remains central to the financing, and much of the downside routes back to Meta through the guarantee. This is why “hidden debt” is imprecise. The obligations are disclosed; a diligent analyst can find them in the footnotes. The sharper description is economically debt-like commitments presented outside the conventional debt total.

The Bank for International Settlements gave this a name in its March 2026 Quarterly Review: shadow borrowing. An SPV raises debt privately; the hyperscaler commits to a long-term operating lease that shows up as an operating expense rather than a balance-sheet liability. Economically it has taken on a long-term financial commitment; on paper it has merely agreed to rent. The BIS also flagged the second-order risk: banks backstop these vehicles with funding lines, embedded guarantees can activate unexpectedly, and private credit can pull back exactly when conditions deteriorate — turning an off-balance-sheet convenience into a shock-transmission channel.¹⁰¹¹

Why this isn't Enron — and the part that actually rhymes

The Enron comparison is irresistible and mostly lazy, but it points at something real if you aim it correctly.

Gil Luria of D.A. Davidson put it best: “Enron's crime wasn't having special purpose vehicles. Enron's crime was hiding them.” Enron used entities controlled by its own CFO to fabricate earnings, park losses, and self-deal — none of it disclosed. Today's structures comply with GAAP and IFRS, are disclosed in filings, and bring in genuinely independent capital that takes genuine risk.¹² Even the loudest skeptics grant the distinction. Michael Burry, who has been hammering the sector, doesn't reach for Enron — he reaches for Cisco circa 1999: a real company selling real picks and shovels into a boom whose demand assumptions outran the cash. That's a far more precise fear. It isn't “the books are fake.” It's “the demand might be.”¹³

The circular-financing loop — and a careful fact-check

Now the part worth getting exactly right, because the intuition is close but the precise wording matters.

The simplified worry looks like this: a chipmaker invests in an AI developer; the developer spends the capital on computing capacity built around that chipmaker's hardware; the chipmaker books hardware revenue; the growth props up valuations across the network; and the higher valuations make the next raise easier. Cash leaves one company as an investment and returns as another's revenue.

On the marquee example: in September 2025, Nvidia and OpenAI announced a nonbinding plan under which Nvidia intended to invest up to $100 billion as OpenAI deployed at least 10 gigawatts of Nvidia systems, with the investment going in as cash and most of it expected to come back to Nvidia through chip purchases or leases.¹² That $100 billion was never completed as announced — by early 2026, Nvidia's CFO acknowledged there was still no definitive agreement, and CEO Jensen Huang said the full figure was “probably not in the cards.”³ What Nvidia actually did was take a $30 billion stake as part of OpenAI's $110 billion round, which closed on February 27, 2026 (alongside roughly $50 billion from Amazon and $30 billion from SoftBank, valuing OpenAI near $730 billion).⁴⁵

So yes: Nvidia is putting real cash into OpenAI while OpenAI commits to infrastructure built around Nvidia hardware. But the public materials do not establish that OpenAI is contractually required to route Nvidia's exact investment dollars back through chip purchases. Money is fungible, and much of OpenAI's compute is procured through cloud and infrastructure partners rather than a direct GPU buy. The precise phrasing is: this is economically circular without necessarily being contractually circular. That distinction keeps you from the overreach of saying Nvidia is booking a sale to itself — while preserving the real question, which is how much end-market demand exists independent of vendor-supplied capital.

Why route it as cash-for-equity rather than chips-for-shares?

Because the two transactions do different jobs. OpenAI doesn't need chips alone; it needs complete operating capacity — land, power, cooling, networking, buildings, cloud services — none of which Nvidia satisfies by shipping GPUs. Cash also lets the investment and the procurement be priced separately, with their own valuations, warranties, and delivery terms. A contribution of chips for equity wouldn't eliminate the accounting questions; it would create a non-cash transaction requiring both the hardware and the equity to be valued, making it harder to prove the two prices were negotiated at arm's length. And Nvidia genuinely wants both things: an investment return if OpenAI appreciates, and commercial revenue from supplying the gear. The cash route is not irrational. It simply makes one question load-bearing: would OpenAI, and the providers serving it, buy the same Nvidia capacity on the same schedule if Nvidia weren't also the one supplying the capital?

This isn't confined to one deal. Microsoft has put roughly $13 billion into OpenAI, much of which flows back as Azure spend; Oracle, CoreWeave, AMD, and SoftBank all sit somewhere in the same ring of investment, offtake, and supply.¹⁴ The cluster is bound together by well over a trillion dollars in committed future spending, with OpenAI in the keystone position — the one entity whose failure would simultaneously impair the demand assumptions, the revenue forecasts, and the collateral values of nearly everyone else.

When there should be concern — follow who takes the loss

No single item here proves anything. Concern should rise as several appear together. The organizing question underneath all of them is the same: who actually loses money if the optimistic case doesn't arrive?

1. The sponsor claims risk was transferred while guaranteeing the outcome. A minority stake is not persuasive if leases, purchase commitments, completion guarantees, or residual-value guarantees leave the sponsor absorbing most of the downside. The test: who loses if the facility is late, obsolete, underutilized, or worth less than its debt?
2. Commitments grow faster than independently generated cash flow. Infrastructure is ultimately justified by paying customers outside the financing network. If capacity purchases and leases outrun cash from unrelated customers, the ecosystem may be financing anticipated demand rather than demonstrated demand.
3. The debt leans on one tenant or one interconnected group. A data center with a single AI customer isn't diversified infrastructure; it's a concentrated credit exposure on a specialized asset — worse when that customer depends on financing from the very companies it's buying from.
4. The collateral's useful life is shorter than the debt. Buildings last decades; GPUs do not. Debt structured around optimistic residual values gets fragile when the next chip generation lands. Long-dated financing does not turn short-lived technology into long-lived collateral.
5. Revenue depends materially on customers the vendor is funding. Ask the supplier to disclose revenue from companies it holds equity in, revenue backed by vendor financing or guarantees, and how much of the order book ultimately rests on outside end-user demand. Fast growth means less when the seller is financing the buyer.
6. Maximum exposure dwarfs reported debt. When guarantees, uncommenced leases, and funding commitments are large relative to on-balance-sheet debt, conventional leverage ratios are simply incomplete. Build an adjusted view.
7. Complexity rises while disclosure quality falls. Complexity can be necessary; opacity is a choice. If you can't answer who owns the asset, who borrowed, who controls it, who provides its revenue, who guarantees the shortfall, and which parties wear multiple hats — that's the real Enron line. Not the structure. The inability to see it.

What this does to valuation

If you're pricing these names — or holding them in an index and pretending you aren't — the off-balance-sheet reality moves four numbers.

Enterprise value and true net debt. You have to fold the SPVs and the deferred lease obligations back in to get real leverage. Headline net-debt and Debt/EBITDA are understated wherever the borrowing has been ring-fenced, so the equity looks cheaper on reported figures than on rebuilt ones. This is the same discipline that catches hidden vendor debt on a private company's books — only the vehicle is a Wall Street SPV instead of an unrecorded payable.

Quality of revenue. Circular, vendor-financed revenue does not deserve the multiple you'd assign organic, third-party revenue. A disciplined analyst haircuts the round-tripped portion before capitalizing growth. Peak enthusiasm does the opposite.

Returns on capital. Keeping the asset and its financing off the books mechanically flatters ROIC and ROA — you're crediting a return against a smaller reported base than the company actually commands. Rebuild the base and the return compresses. The same logic applies to depreciation: every year a company stretches the assumed useful life of fast-aging hardware, it flatters current earnings and defers the write-down. Watch the useful-life assumption closely.

Cost of capital and terminal value. Private-credit spreads and embedded leverage mean the enterprise is riskier than its investment-grade optics imply, so the discount rate should be higher — and terminal value, which rests almost entirely on demand flowing through the keystone, should be stressed, not extrapolated. It's the same question you face valuing any business whose number leans on one set of optimistic assumptions: change the assumption and watch what's left.

Put it together and the conclusion isn't “it's a fraud, short it.” It's more disciplined and more useful: the structures are legal and the debt is disclosed, but the reported financials systematically make leverage look lower, returns look higher, and demand look more organic than a rebuilt model would show. Whether the equity is mispriced turns entirely on whether the end demand is real. That's the Cisco question, not the Enron question — and it's the one worth watching.

Distributing risk does not make it disappear. Follow the cash, follow the guarantees, follow the customer concentration, then model who eats the loss when the optimistic case doesn't show. If all three paths lead back to the same handful of companies, the risk isn't as dispersed as the legal structure makes it look. It's just been renamed.