Five of the largest technology companies on earth are carrying about $1.65 trillion in off-balance-sheet obligations tied to AI infrastructure — more than the $1.35 trillion of debt they actually report. According to the Nikkei study behind the figure, that off-books total has risen roughly eightfold in four years.

The Enron comparison in the headlines is doing a lot of rhetorical work, and it’s worth separating what’s genuinely alarming here from what isn’t.

The structure

Alphabet, Microsoft, Amazon, Meta and Oracle are financing the AI buildout — chips, servers, power, buildings — by packaging the debt into separate legal entities, frequently joint ventures. The obligations sit in those vehicles rather than in the parent’s main accounts.

The individual numbers are the part that makes the aggregate real:

Meta carries roughly $420 billion off balance sheet, close to triple its reported debt. Its Hyperion data centre in Louisiana holds about $27 billion of debt in a separate structure — with Meta as the sole tenant.

Oracle has roughly $260 billion in future lease commitments, and its off-books total has grown around thirtyfold in four years.

Nvidia sits adjacent to the same machine with about $119 billion in purchase obligations.

Sole tenant is the detail worth pausing on. When a company is the only occupant of a facility financed through a vehicle it doesn’t consolidate, the economic exposure and the accounting treatment have drifted some distance apart. The building exists because Meta needs it, is used only by Meta, and would have little alternative use — but the debt is somewhere else.

Where the Enron analogy holds, and where it breaks

The most useful thing said about this came from analyst Gil Luria, quoted by Bloomberg Law: “Enron’s crime wasn’t having special purpose vehicles… Enron’s crime was hiding them.”

That distinction matters enormously, and it cuts against the framing of the story it appears in. These structures are disclosed. They appear in filings. Analysts at S&P, Morgan Stanley and Moody’s have flagged the risks publicly. A researcher was able to total them up — which is precisely what nobody could do with Enron. Post-Enron accounting rules exist and are, on the available evidence, working as intended.

So this is not fraud, and calling it the trick that toppled Enron is imprecise. What it is, is a set of legal, disclosed structures whose combined scale is large enough that the disclosure being technically available doesn’t mean the risk is widely understood. Those are different problems. The second one is still a real problem.

The honest version: an investor reading a headline balance sheet gets a materially incomplete picture of these companies’ AI commitments, and has to do meaningful work in the footnotes to assemble the whole thing. Legal, disclosed, and quietly enormous is a defensible description. Criminal is not.

The mechanism that makes this fragile

Here is the part that deserves the most attention, because it’s a timing problem rather than a transparency one.

When a data centre goes live, the lease moves onto the balance sheet. Immediately. The obligations currently distributed across joint ventures don’t stay out there indefinitely — they arrive, in sequence, as facilities come into service.

That means the current picture is not a steady state. It’s a queue. The next few years will see these commitments migrate onto reported balance sheets on a schedule determined by construction timelines rather than by business performance. If AI revenue grows into the capacity, that migration is unremarkable — it’s just how financed capital expenditure works. If demand undershoots, companies inherit large fixed obligations for infrastructure that isn’t earning, at exactly the moment their reported leverage jumps.

And the exposure doesn’t stop at the tech companies. The debt sits with lenders and insurers. That’s the transmission channel by which a shortfall in AI demand stops being a technology-sector story and becomes a credit story.

Why this rhymes with everything else in AI finance

This fits a pattern we’ve been tracking. AI clouds have started borrowing against their GPUs as collateral, treating compute as a financeable asset class the way airlines treat aircraft. Now the hyperscalers are financing the buildings and power through structures that keep the leverage off the headline accounts.

Both are rational responses to the same pressure: the capital required for AI infrastructure has outrun what operating cash flow comfortably funds, so the industry has turned to increasingly sophisticated financial engineering to bridge the gap. That’s what mature industries do. It is also, historically, what industries do shortly before discovering whether the demand they built for was real.

The read

Strip out the Enron framing and the substance is still significant: five companies have accumulated $1.65 trillion in AI-related obligations outside their reported debt, up roughly eightfold in four years, and those obligations will land on balance sheets as facilities come online rather than when it’s convenient.

The right question isn’t whether the accounting is legal — it is. It’s whether the AI revenue arrives on a schedule that matches the lease schedule. Everything else is bookkeeping. That single question decides whether this is prudent capital structuring or the most expensive bet in corporate history, and nobody currently knows the answer.

Reporting on published analysis as covered on 22 July 2026, based on a Nikkei study and commentary reported by Bloomberg Law. The structures described are disclosed and legal; no allegation of wrongdoing is made or implied. Figures are third-party estimates. Not investment advice.