explainx / blog
Nikkei Asia found Alphabet, Microsoft, Amazon, Meta, and Oracle carry $1.65T in off-balance-sheet AI debt. Is it Enron 2.0 or normal SPV accounting?
Jul 15, 2026
The Bank for International Settlements warns the AI capex wave is outgrowing internal cash — private credit to AI-related borrowers hit $200B+ with spreads matching non-AI loans while equity prices imply megafuture returns. explainx.ai breaks down Bulletin No 120 and Hacker News reactions on too-big-to-fail and missing downside scenarios.
Jul 16, 2026
Anthropic filed a confidential S-1 on June 1 and closed Series H at $965B on May 28. By July 15, bankers were lining up institutional meetings — reports point to a possible October 2026 listing, but Anthropic has not confirmed a date. explainx.ai explains what changes for Claude Code, Fable, API buyers, and what an IPO does not guarantee.
Jul 12, 2026
Thomas Salopek's team tested regime-classifying AI agents across two decades of history — all eight beat 60/40 risk-adjusted. JPMorgan warns against in-sample confidence. explainx.ai reads Polymarket hype vs backtest reality.
Nikkei Asia went looking for how much of the AI data-center boom is actually financed with debt investors can't see on a balance sheet. The answer: an estimated $1.65 trillion, spread across Alphabet, Microsoft, Amazon, Meta, and Oracle — more than the $1.35 trillion these same five companies officially reported as debt for the most recent quarter, according to reporting by Futurism's Victor Tangermann.
That's not a rounding error. It means the debt investors can't easily see from a headline balance-sheet line is larger than the debt they can see. Commentators have reached for an obvious historical comparison: Enron, which collapsed in 2001 after hiding debt in shell companies. Six months after the Bank for International Settlements warned that AI capex was outgrowing hyperscaler cash flow and pushing companies toward debt, this is what that shift looks like in dollar terms — and it's bigger than the BIS bulletin's own private-credit estimate.
| Question | Answer |
|---|---|
| How much off-balance-sheet debt is there? | An estimated $1.65 trillion across Alphabet, Microsoft, Amazon, Meta, and Oracle, per Nikkei Asia via Futurism |
| How does that compare to reported debt? | It's more than the $1.35 trillion these five companies officially reported on balance sheets last quarter |
| Which company has the most disclosed individually? | Meta, at roughly $420 billion — other companies' individual shares aren't fully public |
| Is this illegal? | No. It's disclosed via special purpose vehicles (SPVs) and footnotes to quarterly filings, under standard accounting rules |
| Is this really like Enron? | The mechanism (off-balance-sheet vehicles) rhymes; Enron's actual offense was fraud — hiding debt from auditors and fabricating earnings, not just using SPVs |
| Does this affect regular investors/retirees? | Possibly indirectly, via private credit held by insurers and pension-adjacent funds — but the scale of that specific exposure isn't fully public |
Here's the split Nikkei reported, as summarized by Futurism:
| Category | Amount |
|---|---|
| Debt officially reported on balance sheets (5 companies, most recent quarter) | $1.35 trillion |
| Estimated off-balance-sheet debt (same 5 companies) | $1.65 trillion |
| Meta's individual off-balance-sheet debt | ~$420 billion |
| Companies covered | Alphabet, Microsoft, Amazon, Meta, Oracle |
Two things are worth sitting with. First, the off-balance-sheet figure is larger than the on-balance-sheet figure — this isn't a footnote-sized add-on, it's the majority of the companies' real debt load once you count both. Second, a full per-company breakdown beyond Meta's $420 billion isn't publicly available yet. We're not going to guess at Alphabet's, Amazon's, Microsoft's, or Oracle's individual shares — Nikkei's reporting (paywalled) and each company's own 10-K footnotes are the primary sources for that, and they should be checked directly by anyone making decisions based on these numbers.
Four of the five companies were scheduled to report Q2 2026 earnings within days to weeks of this story breaking, per the article — so the specific figures here are a snapshot, not a permanent scoreboard.
The mechanism is a special purpose vehicle (SPV): a legally distinct subsidiary or joint venture that borrows money — often to build and own a data center — while the parent company signs a take-or-pay contract to lease capacity from it for years. The SPV holds the debt. The parent holds an operating obligation, which shows up differently (and often more favorably) in headline debt ratios than a straight loan would.
This isn't a new trick invented for AI. Airlines have used SPVs to finance aircraft for decades; utilities use them for power plants. What's new is the scale — $1.65 trillion of it, concentrated in five companies racing to build data centers faster than their operating cash flow can fund them.
Crucially, this is disclosed. It shows up in annotations to quarterly financial statements — footnotes in 10-K filings that describe lease commitments, guarantees, and variable-interest entities. That's the legitimate, GAAP-compliant part. The problem isn't concealment from regulators; it's that a retail investor skimming a press release, a stock app summary, or a headline "debt" figure will never see it. Someone who reads the actual footnotes will.
Enron collapsed in 2001 after using entities like the now-infamous "Raptor" and "Chewco" partnerships to move debt and losses off its books, inflate reported earnings, and mislead its own auditors (Arthur Andersen) and the SEC about what those entities actually were and who controlled them. That was fraud — not just an accounting structure, but active deception about the structure's purpose and ownership.
The mechanism the five AI giants are using — SPVs and take-or-pay data-center leases — is structurally similar in the sense that debt sits outside the primary balance sheet. That's genuinely where the Enron comparison earns its keep, and why commentators reach for it.
Where it breaks down: nobody is alleging these companies are lying to auditors about who controls the SPVs, or fabricating revenue to cover the gap. As Tom Selling, a technical accounting consultant, put it via Bloomberg reporting cited by Nikkei:
"The accounting treatment itself is in fashion... what if [it turns into a] house of cards and was propping itself up with this accounting treatment? To me, that's the risk."
Note the framing: the risk Selling names is the accounting treatment itself becoming load-bearing for a shaky structure — not that the treatment is currently fraudulent. That's a meaningfully different claim than "this is Enron." Several Hacker News commenters made the same point when this story circulated: labeling standard GAAP footnote disclosure as "hiding" debt is doing a lot of work in the headline, since Enron's downfall required proving intent to deceive, not just an accounting structure that happens to look similar decades later.
The fair synthesis: this is not fraud. It is opacity that functions similarly to concealment for anyone who doesn't read past the headline number — and that's a real problem for retail investors and financial journalism, even if it's not a legal one.
This is where the debate gets more interesting than "fraud or not fraud." Two threads matter.
Private credit and insurers. Much of this off-balance-sheet debt is funded through private credit — non-bank lenders holding loans to maturity rather than trading them. The IMF's Global Financial Stability Report has separately flagged a related concern: private-equity-controlled insurers are growing their exposure to less-liquid private-credit assets. Insurers hold reserves against life-insurance and annuity products that ordinary people depend on. If AI-linked private credit underperforms and that exposure is concentrated enough, the losses could, in principle, reach further into the financial system than a simple "tech stock correction" would. The open question — and it is genuinely open — is how concentrated that exposure actually is, which isn't fully disclosed yet.
S&P 500 concentration. AI-exposed megacaps dominate the top of the S&P 500 — Nvidia alone is around 7.5% of the index. A sharp correction in one or two of these names could meaningfully move index-tracking retirement accounts, which is how this reaches ordinary savers even without touching private credit at all. But commenters pushed back hard on the more alarmist version of this argument: the "top 10" of the S&P 500 isn't synonymous with "AI stocks" — it includes companies with diversified revenue that wouldn't necessarily fall in lockstep even if AI capex sentiment soured. Concentration raises correlation risk; it doesn't guarantee a 50% wipeout.
Both of these are the same shape of question the BIS bulletin raised in January: leverage "does not disappear by being out of sight." The Nikkei numbers are the first time that abstract warning has come with a concrete, company-level dollar figure attached.
Nikkei's reporting notes these companies aren't only borrowing off-balance-sheet — they're also issuing new shares to help fund the buildout. That matters because it's a second lever pulling in the same direction as the debt story: equity dilution. Existing shareholders get a smaller slice of future earnings even as the company's total capital commitments grow. It's the flip side of the equity-vs-debt pricing schism BIS flagged — debt markets price AI risk conservatively while equity markets price in transformational upside, and now companies are tapping both channels simultaneously to close the capex gap.
This all lands in a market where an Anthropic IPO is being scheduled with bankers and where JPMorgan's own AI-agent backtests came with explicit caveats about not mistaking in-sample research for live performance. Financing structure is quietly becoming as important a story in AI as model capability — see also how Nvidia's revenue-share arrangements with AI startups extend the same "who really carries the risk" question down the supply chain.
A few things worth being direct about.
This is a secondary-source chain. Futurism reported on Nikkei Asia's investigation; Nikkei Asia is paywalled, so we're citing it as the underlying source without linking to a paywall. The actual primary sources are each company's 10-K footnotes and SEC filings. Anyone making an investment decision based on this story should go read those footnotes directly, not rely on this post or Futurism's summary of Nikkei's summary of the filings.
The "bubble" framing is contested. This story is compatible with a bear case (data centers become stranded assets if AI revenue disappoints) and a bull case (Jevons paradox — cheaper compute historically drives more total usage and revenue, not less, as happened with coal and photovoltaics). Both arguments have credible backers. This post is not taking a side on which one wins; it's explaining what the debt structure actually is and isn't.
This is not investment advice. Nothing here should be read as a recommendation to buy, sell, or avoid any of the five companies named.
Nikkei Asia found that five AI giants carry more debt off their balance sheets ($1.65 trillion) than they report on them ($1.35 trillion), funneled through SPVs and take-or-pay data-center contracts. That's disclosed in footnotes, not hidden from regulators — so calling it "Enron" oversells the legal reality, since Enron's crime was fraud, not merely using off-balance-sheet structures. The legitimate worry is opacity: most people reading a headline will never see the $1.65 trillion, and the exposure may reach pensions and 401(k)s indirectly through private credit and S&P 500 concentration. Whether this resolves as a manageable financing footnote or the AI industry's first real credit event depends on earnings that are due within weeks of this story, not on the accounting structure itself.
Sources: Futurism, Victor Tangermann, "Tech Giants Have Hidden Trillions in Off-the-Books Debt, Investigation Finds" (July 22, 2026) · Nikkei Asia investigation (original reporting, paywalled) · Bloomberg, quoting Tom Selling on accounting treatment risk · Hacker News discussion thread, July 2026
This post reflects information available as of July 23, 2026. Four of the five companies named were due to report Q2 2026 earnings within days to weeks of this story — check each company's latest 10-K and quarterly filings directly before drawing conclusions, and treat this as reporting context, not financial advice.
