Read the balance sheets of Alphabet, Microsoft, Amazon, Meta, and Oracle this earnings season, and AI spending looks aggressive but contained. Read the footnotes, and a second company appears, one carrying more debt than the one on the cover page.
That is the finding of a Nikkei Asia investigation published in late July: the five companies have amassed an estimated $1.65 trillion in AI-related obligations that do not show up as debt, compared with roughly $1.35 trillion that do. The hidden figure has grown roughly eightfold since 2022. It is now the larger of the two numbers, and most investors have never seen it.
How debt disappears legally
There is no fraud here, and that is the uncomfortable part. Under current accounting rules, a company does not have to record a liability for a data center lease that hasn't started, or a GPU order that hasn't been delivered. Those obligations appear in the notes attached to a 10-Q rather than on the balance sheet itself. Once the facility goes live or the hardware ships, the number moves onto the books all at once.
That structure is exactly what let Meta, Oracle, and their peers finance a trillion-dollar infrastructure race without it registering as leverage in the metrics most investors actually check. Analyst Gil Luria, speaking to Bloomberg Law about the parallel, put it plainly: the accounting scandal that took down Enron wasn't the use of special purpose vehicles. It was hiding them. What's happening now is the same mechanism, done in the open, inside the rules.
Meta and Oracle: the two numbers we actually have
Only two of the five companies have individually sized hidden-debt figures in the public record so far, and both are instructive.
Meta's off-balance-sheet obligations are estimated at $420 billion, close to three times its reported debt of roughly $140 billion. A large share runs through a joint venture with Blue Owl Capital that is financing the company's Hyperion data center campus in Louisiana, in which Meta holds only a 20% stake but has reportedly guaranteed to cover investor losses if the facility is later scaled back.
Oracle's hidden debt is smaller in absolute terms at roughly $273 billion, but its growth rate is the sharper story: a more than thirtyfold increase in four years, funded through project financing tied to AI buildouts in Texas and Wisconsin, including its role in the Stargate venture with OpenAI. S&P Global has already acted on that exposure, cutting Oracle's rating to one notch above junk on July 9, citing the AI-driven debt load directly.
Alphabet, Microsoft, and Amazon have not had comparable individual estimates confirmed in public reporting. That is worth sitting with rather than glossing over. Alphabet's most recent quarterly filing discloses $40.7 billion in future funding commitments to off-balance-sheet vehicles, a real and specific number, but a narrower category than the full hidden-debt estimate Nikkei built for Meta and Oracle. The other three companies' true exposure may simply not be sized yet in a form the market can price. In a sector where the debt has grown eightfold in four years, an unpriced gap is not a footnote. It is a risk.
Why regulators got here before the headline number did
The Bank for International Settlements flagged this structural gap before Nikkei attached dollar figures to it, describing hyperscaler financing arrangements as a form of shadow borrowing that is economically equivalent to debt even though it sits outside the usual disclosure channels. Moody's separately warned that rising lease commitments were weighing on credit quality across the sector, even in cases where headline debt looked manageable.
The mechanism both institutions point to is the same one driving the Nikkei numbers: companies are locking in long-term data center leases and GPU supply contracts instead of issuing conventional debt, because the accounting treatment is more forgiving. BIS also noted that where public bond issuance does happen, it tends to run long, typically more than five years, while the lease and joint-venture layer underneath it is structured around multi-year capacity commitments that don't map cleanly onto a single bond-style maturity.
The mismatch underneath the number
The part of this story that deserves more attention than it has gotten is timing risk, not just size. AI accelerator hardware is now on an 18- to 36-month replacement cycle, driven by a new generation of chips roughly every 18 months, according to Electronics Weekly's reporting on the same Nikkei data. The financing built around that hardware, by contrast, often stretches far longer, sometimes a decade or more when leases and infrastructure debt are included.
That gap matters because the collateral backing much of this financing ages out of relevance long before the obligations tied to it come due. A server bought against a ten-year lease may be functionally obsolete in three. Standard credit models, built around comparing debt levels to cash flow, were not designed to capture that kind of asset-duration mismatch, because it barely existed at this scale before the AI buildout.
What to actually watch
None of this means the AI infrastructure bet is wrong. Cloud demand is real, and companies like Amazon Web Services have pushed back hard on the idea that this spending is speculative. What it means is that the balance sheet alone is no longer a reliable way to measure how leveraged these companies actually are.
The next signal to watch is not another headline debt number. It's whether credit analysts and short sellers start pricing lease and supply commitments alongside bonds as a matter of routine, and whether Alphabet, Microsoft, and Amazon end up disclosing the kind of company-level breakdown Meta and Oracle already have, voluntarily or otherwise.