On 10 August 2026, Nvidia stood on a stage with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR and announced a plan to mobilise over $500 billion in third-party capital for AI infrastructure. Jensen Huang called it a milestone: compute, he said, had become "productive, investable infrastructure." Six chief executives gave a live joint television interview the same day to reinforce the message. The optics were immaculate.
Three days later, on 13 August, the Telegraph ran a piece on Michael Burry accusing the same company of "shades of Enron." The juxtaposition is instructive, though not for the reason the headline implies. Burry is not alleging fraud, and has said so directly; his own clarification, after Nvidia's memo pushed back, was that he is not claiming Nvidia is Enron, that "it is clearly Cisco." The comparison he is actually making is to 1999 and 2000: overbuilt supply meeting demand projections that never materialised, dressed in extended depreciation schedules that flatter near-term earnings. The distinction matters, because it shifts the question away from whether anyone is lying and towards whether anyone can actually see the structure clearly enough to know if it is sound.
That second question has already been answered once this year, by an institution with no interest in AI industry theatre.
In March 2026, three BIS economists, Egemen Eren, Ingomar Krohn and Karamfil Todorov, published a paper on precisely this financing pattern and gave it a name. Hyperscalers' off-balance-sheet vehicles, they wrote, "amount to 'shadow borrowing': obligations that are economically akin to debt but largely reside outside corporate balance sheets." The paper names the hyperscalers using these structures directly: Amazon, Alphabet, Microsoft, Meta and Oracle among them. Banks, the authors note, support these vehicles with funding lines, which they identify as a potential new shock-transmission channel, whether through refinancing pressure at the vehicle level, procyclical shifts in private credit appetite, or guarantees embedded in the structures that could activate unexpectedly.
That is not a rhetorical flourish from a Substack account. It is the Bank for International Settlements, in its own quarterly review, warning that the visible financial metrics for hyperscalers may not capture their full exposure, and that investors relying on traditional balance-sheet ratios could be underestimating actual leverage.
Nvidia sits adjacent to this pattern rather than inside it, and the distinction is worth holding onto. In its November 2025 memo to Wall Street analysts, the company stated plainly that it does not use special purpose entities and does not rely on vendor financing arrangements, drawing an explicit contrast with Enron and Lucent. That claim has not been seriously disputed even by Nvidia's critics; Burry's own clarification, that the fair comparison is Cisco's 1999-2000 overbuild rather than Enron's concealed fraud, effectively concedes the point. The SPVs the BIS is describing sit on the hyperscaler side of the transaction, the Amazons and Microsofts and Oracles building the data centres, not on Nvidia's own books. Monday's $500 billion arrangement extends that same pattern one step further: third-party financing platforms, independent of Nvidia's balance sheet, built specifically so hyperscalers and neoclouds can buy more of Nvidia's hardware. The chips are the collateral underneath the shadow borrowing the BIS flagged; Nvidia is the beneficiary of the demand it finances, not the entity doing the off-balance-sheet borrowing itself.
Monday's $500 billion deal is not a new financing instrument. It is the existing shadow-borrowing channel, already flagged by the BIS five months earlier, scaled up and given a press conference. Four of the six institutions on stage with Huang, Apollo, BlackRock, Blackstone and Goldman Sachs, are among the same firms private credit research already identifies as the principal originators of data centre debt, both through off-balance-sheet SPV transactions and direct lending. The MOUs formalise a pattern that was already the dominant financing route; they do not depart from it.
Huang did not wait to be asked about circularity. In the announcement itself, he wrote that the arrangement was "designed to address" the concern directly, stating the capital involved "is not Nvidia revenue" and that underlying demand is real. Some analysts took the point: Bank of America called it a pivot away from the vendor-financing structure that had drawn the circularity criticism in the first place. Others were unmoved. Morgan Stanley's Joseph Moore noted that the debate over how much leverage the AI ecosystem can absorb persists regardless of whose balance sheet it sits on; Mizuho's traders wrote that the deal does not answer how much end-user demand actually underlies the spending. The disagreement is not about the facts of the structure. Both sides can see the same MOUs. It is about what those facts imply, which is precisely the kind of dispute a market pricing mechanism, rather than a press release, is supposed to resolve.
The gap between institutional claim and institutional action runs through the regulatory record as well as the financing structure.
In November 2025, a group of US senators wrote to Treasury Secretary and FSOC Chair Scott Bessent asking the Financial Stability Oversight Council to convene a working group on AI valuation risk, warning that hidden leverage, circular financing and off-balance-sheet commitments could transmit stress across financial markets and nonbank counterparties. FSOC exists for exactly this purpose; it was created after 2008 specifically to catch cross-market contagion before it compounds.
In January 2026, Senator Elizabeth Warren and colleagues wrote again, more pointedly, naming the more than $1 trillion in debt projected for AI infrastructure and asking Bessent to confirm, by 13 February 2026, that FSOC had launched a formal investigation. As of the most recent reporting available for this piece, no public confirmation of that investigation exists.
By June, Warren and Senator Richard Blumenthal stopped waiting on FSOC and moved to legislate around it. The AI Bubble Transparency Act, introduced 11 June 2026, would direct the Office of Financial Research to compel financial institutions to disclose their debt and equity exposure to chipmakers, data centres, cloud providers and hyperscalers, reporting to Congress within a year. Warren's own framing of why the bill is necessary doubles as a description of the gap this piece is tracing: AI and Big Tech companies, she said, are "increasingly reliant on shadowy forms of debt and balance sheet magic," and the bill exists to give regulators the information to catch risks early rather than after the fact. A senator writing new legislation to compel the disclosure FSOC could plausibly have already been gathering is itself evidence of how little confidence there is that the existing mandate is being discharged.
It is not that FSOC did nothing in this period. It ran a four-part AI Innovation Series from March to May 2026, convening banks and AI developers on strategy, value generation, cybersecurity and financial stability. But the series concluded not with a leverage warning, rather with FSOC's own Deputy Assistant Secretary, Christina Skinner, arguing publicly that regulatory caution is itself the risk: frameworks built for a different era, she said, can produce "a reluctance to innovate, and that reluctance in itself can become a source of financial stability risk." Treasury Secretary Bessent has separately been developing a FINRA-style oversight body, reported in July, but its remit is vetting frontier AI models for cybersecurity risk before release, not tracking debt exposure across the financial system. Two live regulatory tracks, both getting real institutional energy in the same window; neither is the debt-and-leverage investigation Congress explicitly asked for.
Meanwhile the Federal Reserve's own May 2026 Financial Stability Report found that half of surveyed market contacts now cite AI as a salient risk to the financial system, up from 30 per cent in the fall 2025 survey; private credit specifically saw an even sharper jump, from 22 per cent to 50 per cent, in part because of AI's disruptive pressure on the software borrowers who make up a large share of private credit portfolios. A one-survey jump from a minority concern to a majority one is not a fringe indicator, on a subject regional bank presidents are now publicly split over: New York's John Williams sees strong borrower earnings and no bubble; Kansas City's Jeff Schmid is asking openly whether AI infrastructure has become too big to fail.
An undischarged congressional mandate, a majority of the Fed's own market contacts now naming AI a stability risk, and a headline financing deal proceeding regardless are not three separate stories. They are one story about which institutional claims get load-bearing scrutiny and which get a press release.
The asset-backed lending model this deal extends has historically depended on collateral holding value independent of the borrower's fortunes. Property, aircraft, toll roads: repossess them and there is still a market.
GPUs do not behave this way. They depreciate on a technology curve, not a wear curve. An H100 fleet does not wear out; it becomes commercially uncompetitive the moment the next architecture delivers materially more throughput per watt. Burry's sharpest technical point is that hyperscalers have been extending useful-life assumptions on this hardware from two to three years out to five and six: Microsoft moved from four years to six, Meta adopted a 5.5-year schedule, and Burry estimates the industry-wide effect is roughly $176 billion in understated depreciation between 2026 and 2028, projecting Oracle's reported earnings could be overstated by 26.9 per cent and Meta's by 20.8 per cent by 2028. These are Burry's own estimates, built on his own assumptions about true useful life, not audited figures, and they should be read that way. Nvidia disputes the framing but has not disputed that the extensions occurred.
If a borrower in one of these financing platforms defaults, the lender is left holding compute that is ageing on a curve set by Nvidia's own product roadmap, not by physical wear, into a resale market that is thin by construction, since the primary buyers for last-generation compute are the same hyperscalers who are simultaneously the primary sellers of the demand story propping up the collateral's valuation. That circularity is the mechanism, not a rhetorical description of it.
This is where the piece needs to leave the abstraction behind, because the exposure is not confined to sophisticated capital that priced it knowingly.
Private credit funds, principally Apollo, Blackstone, Blue Owl, PIMCO and BlackRock, now originate most data centre debt. Their own capital is funded upstream by pension funds, insurers and endowments seeking yield. That is the first layer, and it is the layer the current regulatory framework assumes is sophisticated enough to underwrite its own risk.
The second layer is where "contained" stops being a safe word. Federal Reserve research already shows that up to a quarter of bank loans to non-bank financial institutions now go to private credit firms, up from one per cent in 2013, and that major life insurers are carrying close to $1 trillion in private credit exposure. Banks reporting light direct AI-adjacent exposure, around 0.8 per cent of total assets by one Chicago Fed estimate, are nonetheless indirectly exposed through their lending to the funds doing the actual lending. This is structurally the same daisy-chain that turned an apparently contained subprime problem into a systemic one in 2008: no single institution looks dangerously exposed, and the aggregate exposure is exactly the thing nobody with a full view is required to report.
The third layer is the one worth leading with for a general reader, because it is the one that makes this personal rather than theoretical. The private placements financing these deals, 144A securities, cannot be bought directly by retail investors, but they sit inside the mutual funds and target-date funds that make up more than 40 per cent of 401k holdings, because any fund benchmarked against the standard global bond index will hold them by design. Ordinary retirement savers are already financing part of this build-out, through the most conservative-looking part of their portfolios, without having been told. In March 2026, the US Department of Labor proposed opening 401k accounts to private credit directly, which would remove even that layer of indirection.
Institutions with the resources for manager-level due diligence are told the shadow-borrowing structure is fine because it is professionally managed. The exposure is migrating, in parallel, into the retirement accounts of people who were never offered that due diligence and never asked whether they wanted the position.
That is the gap between the claim made on stage on 10 August and the operational reality the Bank for International Settlements, the Federal Reserve's own survey data, and an unconfirmed congressional mandate all point to underneath it.
BIS Quarterly Review, March 2026, Box A: Eren, Krohn, Todorov, "Financing the AI infrastructure boom: on- and off-balance sheet borrowing" — bis.org/publ/qtrpdf/r_qt2603a.pdf
NVIDIA Newsroom, "NVIDIA Partners With Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to Catalyze Over Half a Trillion Dollars," 10 Aug 2026
Federal Reserve Board, Financial Stability Report, May 2026 — federalreserve.gov
Senate Banking Committee (Warren, Blumenthal), "Ahead of Committee Hearing on AI and the American Dream, Warren and Blumenthal Introduce the AI Bubble Transparency Act," 11-12 June 2026, incl. bill text S.4743 — banking.senate.gov
Senate Banking Committee (Warren et al.), letter to FSOC Chair Bessent, 22 Jan 2026
Congressional letter to FSOC Chair Bessent (Foster et al.), 20 Nov 2025
US Treasury, "Treasury Concludes the Artificial Intelligence Innovation Series," press release, 19-24 May 2026 — home.treasury.gov/news/press-releases/sb0540
US Treasury, "Treasury Launches the Artificial Intelligence (AI) Innovation Series," press release, 23 March 2026 — home.treasury.gov/news/press-releases/sb0421
CNBC, "Big Short investor Michael Burry accuses AI hyperscalers of artificially boosting earnings," 11 Nov 2025
CNBC, "Nvidia tries to quiet 'circular financing' accusations. Wall Street is unsure it will," 11 Aug 2026
Bloomberg, "Nvidia Taps Wall Street for $500 Billion Funding Commitment," 10 Aug 2026
The Daily Telegraph, "Big Short investor issues Nvidia warning" (Chris Price), 13 Aug 2026
Yahoo Finance / DataCenterDynamics, coverage of Nvidia's November 2025 Wall Street memo
Bloomberg / Claims Journal, "US Considers Creating Finra-Like Watchdog to Vet Top AI Models," Jul 2026
Federal Reserve Bank of Chicago, "Tail Risk for Banks Posed by Investments in Generative Artificial Intelligence," 2026
Quartz, "AI data center debt might be in your 401k. Here's how it got there," 18 Apr 2026
Chatham House, "Financial regulators need to get ahead of the curve on private credit," Jul 2026
The Sovereign Auditor covers digital sovereignty, cybersecurity governance, and data protection policy -- with a particular focus on Isle of Man jurisdiction and Crown Dependency issues.
Payments via PayPal. Credentials delivered by email. No Substack. No Stripe. No middlemen.
