On July 22, a story dropped that should be making the rounds in every infrastructure team's Slack channel, but isn't getting nearly enough attention from the people who actually build on top of these clouds. Futurism's analysis, quickly confirmed by The Next Web, revealed that the five largest US hyperscalers — Alphabet, Amazon, Meta, Microsoft, and Oracle — are collectively carrying approximately $1.65 trillion in off-balance-sheet debt tied to AI infrastructure. Their publicly reported, on-balance-sheet debt for the same period is $1.35 trillion. That means the debt you can't see in their financial statements is 122 percent of the debt you can.
I've been running hosting infrastructure for a long time, and I've watched the data center financing world evolve from straightforward capital leases to increasingly baroque structures. But this scale is something different. Let me break down what's actually happening, why the accounting works the way it does, and why every engineer and technology leader who depends on these clouds should understand the numbers.
How $1.65 Trillion Disappears from a Balance Sheet
The mechanism isn't complicated, even if the scale is staggering. The big five are financing AI infrastructure through three main structures that GAAP lets them keep off their primary balance sheets:
- Special purpose vehicles and legal subsidiaries. Meta alone has roughly $420 billion in off-balance-sheet obligations — nearly triple its officially reported debt. Meta's Hyperion data center project carries $27 billion in debt packaged into a separate legal entity. That entity is real, the debt is real, but Meta's quarterly filing doesn't reflect it the way a direct obligation would.
- Signed-but-not-commenced leases. Moody's has identified approximately $662 billion in data center lease commitments that have been signed but not yet started. Under GAAP's lease commencement standard, these don't hit the balance sheet until the lease commences, which can be years away. They exist in footnotes, if they appear at all.
- Private credit and off-balance-sheet financing. An estimated $800 billion in data center buildout is being routed through private credit structures specifically designed to avoid public disclosure obligations.
Oracle is perhaps the most dramatic example of acceleration: its off-balance-sheet commitments have grown approximately thirtyfold over four years. The company now carries $260 billion in future lease commitments that won't appear as a liability on the face of its balance sheet until the properties are operational. Nvidia, not itself a cloud operator but the critical supplier, sits on $119 billion in purchase obligations — customers who have committed to buy, but not yet received or paid for, hardware.
The Revenue Gap That Makes This Dangerous
Here's the number that keeps me up at night. The five hyperscalers are on track to spend somewhere between $700 billion and $900 billion on capital expenditures in 2026 alone — a 36 percent increase over 2025. Meanwhile, researchers estimate the annual revenue gap between what the hyperscalers are spending on AI infrastructure and what the entire AI ecosystem generates in actual sales is approximately $600 billion, and it's widening. Microsoft reportedly spent $97 billion on AI over the last four quarters while generating $37 billion in AI-related annual recurring revenue. That's roughly 38 cents of revenue for every dollar spent.
To put the capex intensity in context: the five major hyperscalers are now running capital expenditures at approximately 34 percent of revenue. The peak capital intensity during the 1990s internet buildout — the one that ended in a spectacular collapse — was around 15 percent. We are more than double that level, and the on-balance-sheet figures don't even capture all of it.
None of this means these companies are in immediate trouble. Alphabet, Amazon, Microsoft, and Meta are generating enormous cash flows. They can service this debt. But "can service it now" and "built on a sustainable financial foundation" are different questions, and the gap between them is where the risk for infrastructure operators actually lives.
The Enron Comparison Is Being Made, and It Matters
Gil Luria, an analyst who spoke to Bloomberg Law about these structures, put it bluntly: "Enron's crime wasn't having special purpose vehicles. Enron's crime was hiding them." Tom Selling, a technical accounting consultant, was more pointed: "What if one of these companies was a house of cards and was propping itself up with this accounting treatment? To me, that's the risk."
I want to be careful here, because the Enron comparison gets thrown around recklessly. Enron was committing fraud. There is no allegation of fraud here — these structures are legal and, in most cases, disclosed somewhere in financial filings. The accounting treatment is, as Selling notes, "in fashion." The SPV approach to data center financing has been normalized across an entire industry in a remarkably short period of time.
The relevant part of the Enron parallel isn't criminality. It's what happened when investor sentiment shifted and the off-balance-sheet obligations had to be consolidated. When Enron's SPVs came back onto its balance sheet, the company ceased to exist. The hyperscalers are not Enron. But the structural similarity — massive obligations sitting in legal footnotes while earnings reports emphasize the headline figures — is the kind of thing that can matter very quickly when credit markets tighten or when AI revenue fails to ramp at the pace investors have priced in.
What This Means If You're Building Infrastructure on These Clouds
I want to be direct about the operational implications, because this is where I think the conversation needs to go. Most of the press coverage treats this as a story about investor risk. But for technology leaders and infrastructure teams, the more immediate concern is dependency.
Vendor concentration risk just got reframed. The standard argument for multi-cloud has always been about avoiding lock-in at the feature and API layer. This adds a financial concentration dimension that most teams don't model. If your entire compute stack runs on one hyperscaler, you're not just exposed to their pricing decisions and service outages — you're exposed to their balance sheet trajectory. That's a new variable.
Long-term pricing commitments deserve more scrutiny. When you sign a three-year reserved instance contract or a committed use discount, you're implicitly betting that the provider's pricing structure remains stable. Providers carrying this level of off-balance-sheet debt have an incentive to grow revenue fast enough to service it. What happens to pricing if they need to accelerate that growth? What happens to the contract terms you negotiated if the capital structure underneath them needs restructuring?
The infrastructure buildout itself creates supply dynamics you can exploit. $3 trillion in projected AI data center spending through 2028 means there will be substantial compute capacity available. If the revenue ramp doesn't materialize at the pace the investment implies, you're likely to see spot and preemptible pricing fall significantly. I'm not making infrastructure bets right now that depend on expensive reserved capacity being cheaper than spot will be in 18 months.
Watch the credit markets as a leading indicator. AI-related debt issuance is projected to hit $570 billion in 2026, and bond investors are already pushing back — demanding higher yields as they price in the mismatch between spending and revenue. When credit markets start attaching a premium to AI infrastructure debt, that's the signal that refinancing gets harder. Infrastructure decisions made today should account for what happens to your providers if that premium widens sharply.
The Harder Question Nobody Is Asking
Here's what I keep coming back to: the AI boom has been so fast and so total that nobody has seriously modeled the scenario where the infrastructure spending is right but the revenue timing is wrong by two or three years. The internet analogy is imperfect — this isn't a case of companies building out capacity that nobody will ever use. AI compute will eventually be monetized. The question is whether the financial structures carrying the buildout are durable enough to survive a timing mismatch.
I've watched hosting infrastructure evolve through multiple cycles. The companies that survived the dot-com collapse weren't necessarily the ones with the best technology — they were the ones whose capital structures could absorb a prolonged revenue drought. The hyperscalers are far larger and more diversified than any 2000-era hosting company, but $1.65 trillion in off-balance-sheet obligations is a number that deserves to be named out loud, understood precisely, and factored into every multi-year infrastructure decision you're making.
This isn't a prediction of collapse. It's a recognition that the balance sheet you're looking at isn't the complete picture, and the complete picture is worth knowing.