A Nikkei investigation has found that Alphabet, Microsoft, Amazon, Meta, and Oracle are sitting on an estimated $1.65 trillion in debt that does not appear on their balance sheets. That is more than the $1.35 trillion in debt they officially report. Read that again: the five companies leading the world’s AI infrastructure race owe more money off the books than they do on them.None of this is illegal. It is not even hidden, in the sense of being concealed. It is disclosed, just not in the place most investors, or most headlines, look. The obligations come from long-term contracts for Graphics Processing Units (GPUs) and servers, multi-year data centre leases and joint ventures structured through special purpose entities. Under current accounting rules, none of these count as debt on a company’s primary balance sheet until the underlying facility actually goes live. Until then, they sit in the footnotes.The scale, by company, is striking. Meta’s off balance sheet exposure is roughly $420 billion, nearly triple its officially reported debt of about $140 billion and this figure was confirmed directly by Nikkei’s reporting. Oracle’s hidden obligations have grown more than thirtyfold in four years to roughly $273 billion, driven in large part by its “Stargate” data centre partnership with OpenAI, against reported debt of around $100 billion. Amazon, Microsoft and Alphabet’s figures, roughly $350 billion, $350 billion and $250 billion respectively, are estimated rather than confirmed, based on the pattern established by the two companies whose numbers Nikkei was able to verify directly. All five companies declined to comment on the findings.How the debt disappearsThe mechanism is straightforward once you see it. A company that wants to build a $10 billion data centre has two broad options. It can borrow the money directly and put it on its own balance sheet or it can set up a joint venture in which it holds a minority stake, let a partner – often a private credit fund – finance construction and sign a long term lease or capacity agreement to use the facility once it is built. The second option is economically almost identical to taking on debt. The company is contractually on the hook for years of payments either way. But because it doesn’t hold a majority stake or direct ownership, it doesn’t have to consolidate the liability onto its own books.Also read: Beyond Bigger Models: Why Smarter Scaling Will Define the Future of AIMeta’s data centre project in Louisiana, financed in partnership with private credit firm Blue Owl Capital, follows exactly this pattern. By holding roughly a fifth of the project itself, Meta keeps the bulk of the financing off its primary balance sheet, even though it remains committed to the facility’s output. Multiply that structure across five companies racing to build AI capacity simultaneously and you get a very large number that most standard debt to equity calculations simply never see.The Bank for International Settlements flagged this pattern back in March, describing it as a form of “shadow borrowing” and warning that it creates new, harder to see links between tech companies, private credit funds and insurers, the kind of linkage that can turn a localised problem into a systemic one if conditions turn.Why the timing mattersFour of the five companies named in the investigation report quarterly earnings within the next two weeks. Their reported debt figures, the ones that will actually make headlines, will look manageable, even conservative, next to their cash generation. The $1.65 trillion sitting in the footnotes almost certainly won’t get the same airtime, because it is not the number that earnings day coverage is built to lead with.That gap matters because of what happens next. When a data centre currently under construction becomes operational, its lease doesn’t ease onto the balance sheet gradually, it lands all at once, as a formally recognised liability. If AI demand growth keeps pace with the scale of the infrastructure being built, that transition is a formality. If it doesn’t, if the facilities coming online turn out to be larger than what customers actually need, those assets get marked down and the losses don’t stay contained to the tech companies themselves.Who actually carries the riskThis is the part of the story that tends to get lost. The institutions financing this build out through private credit and project bonds, insurers, pension funds and specialised credit vehicles chasing higher yields, are exposed to the same downside as the tech companies, often without the disclosure requirements that would let their own investors see the concentration building up. A slowdown in AI monetisation wouldn’t just dent a tech company’s share price, it would flow through to the balance sheets of institutions that ordinary savers and policyholders rely on, often without those savers ever knowing the exposure was there.Also read: India Wants to Lead the AI Race, but What About the Workers Behind It?None of this means the AI infrastructure boom is destined to end badly. Demand may well grow into the capacity being built and the companies involved generate enough free cash flow to absorb considerable shocks if it doesn’t. What the Nikkei investigation actually demonstrates is narrower and, in some ways, more important: the accounting rules governing how this generation’s infrastructure race gets reported were not built for commitments moving at this speed or this scale. Investors reading only the headline debt figures this earnings season are, by definition, seeing less than half the picture.Pravin Kaushal is a tech and social entrepreneur and Programme Director (Eastern India) at WHEELS Global Foundation. He writes on AI, economy and geopolitics and can be found on X at @ipravinkaushal.