$1.65T off-balance-sheet: Nikkei's investigation prices the AI debt that shows up nowhere

In plain terms. A Nikkei Asia investigation puts a figure on what credit analysts suspected: Alphabet, Microsoft, Amazon, Meta, and Oracle collectively hold ~$1.65 trillion in AI-related commitments off their balance sheets, compared to $1.35 trillion in official debt. The mechanism is legal, common accounting-wise—but it makes measuring real risk very difficult.

The mechanism: special purpose vehicles (SPVs), legally distinct subsidiaries, structured leasing arrangements. Nothing illegal or new—the oil majors have been doing it for twenty years. What is new: the speed and scale. Meta alone would concentrate ~$420 billion of these commitments—more than a quarter of the total off-balance-sheet identified by the investigation. They cover data centers, compute commitments with third parties, long-term energy contracts.

The credit context aligns. S&P downgraded Oracle to BBB- four days ago (our publications #1359, #1502): estimated FY27 capex of $90-95 B, of which about half of the $638 B in unfulfilled commitments come from OpenAI. The BIS, in its June bulletin, had already mapped the supplier-investor circularity (Nvidia invests in its customers, who buy back chips from it). The Nikkei investigation adds the layer “what you see on the balance sheet is not what exists.”

Tom Selling, a technical accounting consultant quoted by Bloomberg in the Futurism article: “The accounting treatment itself is in fashion. But 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.” Not an accusation of fraud—the reminder that aggregate risk is measured from accounts that only show a third of it.

Under the hood. A typical SPV in the sector: an hyperscaler co-finances a vehicle minoritarily, signs a long lease (15-20 years) with a fixed payment commitment. The future commitment appears in a footnote; the SPV's debt does not appear in the consolidated balance sheet. Similar structures for compute (purchase of GPUs then leaseback) and energy (long-term PPA, withdrawal commitment over 10-20 years). The investigation does not detail SPV by SPV—this precision is missing for the analyst who wants to model the risk.

So what. Three things for a CFO or an investor: (1) the published debt/EBITDA ratios underestimate the cumulative AI-related leverage—credit analysts will recalibrate. (2) A next downgrade would be triggered less by official debt (reasonable) than by the appearance of an incident in an SPV. (3) The margin of error on the thesis “hyperscalers absorb AI capex” is not the one displayed: if a single major SPV stumbles, the contagion effect would be severe. The BIS did not mention “2008 parallels” by chance.