Fund managers have settled on their nightmare for 2026, and they named it out loud: 48 percent of managers surveyed by Bank of America called AI data center borrowing the top systemic credit risk of the year, ahead of government debt (StartupFortune citing the BofA fund manager survey, Jul 22). But the fear is aimed at the wrong ledger. The bonds those managers can see and trade are the smaller, cleaner half of what the AI boom owes. S&P Global counts $225 billion of bonds issued by hyperscalers and related entities like Nvidia so far this year, on pace for roughly $400 billion across 2026 (Fortune, Jul 31). That record number is the visible half. The invisible half runs to about $1.65 trillion.
The contradiction sits in plain sight. The five largest US tech companies carry $1.35 trillion of debt on their balance sheets while owing an estimated $1.65 trillion more in obligations that never appear there, according to a study by Nikkei, mostly long-term purchase deals for chips and servers plus leases on data centers still being built (Fortune, Jul 31). Moody's puts the off-balance-sheet pile at $1.2 trillion, with more than $820 billion of it tied to centers under construction, rent the companies will owe whether or not the machines inside ever earn anything (Fortune, Jul 31). Goldman Sachs analysts count hyperscaler lease commitments of $1.5 trillion, up from about $200 billion five years ago, and note that roughly $1 trillion of that has not even commenced, meaning it shows up nowhere until the buildings open (CNBC, Aug 14).
The trigger this month was Nvidia. On August 10 the company announced memoranda of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to mobilize more than $500 billion of third-party capital into lending platforms for Nvidia's own customers (Steffen credit analysis, Aug 14). Nothing is committed yet; these are non-binding agreements, and Jensen Huang said Nvidia may itself backstop up to $125 billion, a quarter of potential deals (Steffen credit analysis, Aug 14). Huang told CNBC his company's chips are now an "investable infrastructure asset" (CNBC, Aug 14). Translate the phrase and you get the slow pressure underneath the headline: a chipmaker organizing half a trillion dollars of borrowed money so that customers can buy its product, then offering to eat the first quarter of any loss.
Where does that risk actually land? Not on Nvidia, and not on the banks. Steffen traces the flow: the capital comes from insurance general accounts, since Apollo owns the annuity writer Athene and KKR owns Global Atlantic, moves through special purpose vehicles to borrowers like CoreWeave, Crusoe and Nebius, who spend it on Nvidia hardware, and comes back relabeled as investment-grade fixed income (Steffen credit analysis, Aug 14). The template already exists. CoreWeave closed an $8.5 billion loan in March secured by GPUs and customer contracts, rated A3 by Moody's, parked in a bankruptcy-remote vehicle anchored by Blackstone's credit arm (Steffen credit analysis, Aug 14). Meta's Hyperion data center in Louisiana did it bigger: $27 billion of vehicle-level debt rated A+, yielding 6.58 percent at issue, maturing in 2049, with PIMCO reportedly taking about $18 billion and Blue Owl owning 80 percent of the venture, keeping all of it off Meta's books (Steffen credit analysis, Aug 14).
What makes that paper investment grade is not the silicon. It is a take-or-pay contract under which a strong company owes the payments whether or not it uses the capacity. Where the payer is Microsoft or Meta, the substitution works. Where the payer is an AI lab, it fails quietly: OpenAI generates roughly $25 billion of annual revenue against a projected 2026 cash burn near $27 billion and cloud commitments itemized at close to $590 billion, so a contract from it is only as good as its next fundraising round (Steffen credit analysis, Aug 14). Meanwhile investors are already buying protection against Nvidia itself, with credit default swaps on its loans described as spiking (Matterfact weekly, Aug 19).
The history that fits is the one PIMCO strategist Lotfi Karoui reached for himself: adjusted for inflation, he calls this the largest investment cycle since the nineteenth-century railway construction (PIMCO commentary, Aug 11). The railways are the right bounded model. British and American rail manias were financed by visible equity and invisible layers of guaranteed bonds and vendor paper, exactly as AI is financed by visible bonds and invisible leases. The rails got built either way; the people ruined were the ones holding the layered paper when traffic failed to fill the track. What differs this time: the railroads' traffic never arrived, while hyperscalers sit on some of the strongest cash flows in corporate history, which Moody's acknowledges leaves their investment-grade ratings facing no imminent risk (Fortune, Jul 31). The counter-case argues the other way just as hard: debt-funded share of hyperscaler capital spending went from about 9 percent in fiscal 2024 to roughly 32 percent by mid-2026 (Steffen credit analysis, Aug 14), and the marginal borrower in this cycle is no longer Alphabet. It is a neocloud whose collateral depreciates faster than its loan amortizes, because Nvidia ships a new chip architecture roughly every year against debt maturing in 2032 or later (Steffen credit analysis, Aug 14).
The consequences run downhill in a fixed order. First to pay are the yield buyers: insurers and pension funds holding paper rated on the strength of contracts from counterparties who themselves need fresh capital to honor them. Michael Burry has built a short thesis estimating hyperscalers understate depreciation by $176 billion between 2026 and 2028 (Steffen credit analysis, Aug 14); if he is even half right, the collateral behind hundreds of billions in vehicles shrinks while the coupons keep getting paid, and nobody notices until a refinancing fails. Second to pay are the equity holders of the borrowers, since these structures are non-recourse: the vehicle defaults, the GPUs go back to the lenders, and the sponsor walks. Third comes the market plumbing. A borrowed-money AI hedge fund called Situational Awareness already showed the speed available here, falling from $45 billion in assets to about $10 billion on margin calls before Ken Griffin's Citadel bought its listed positions at a discount (CNBC, Aug 14). And the whole wave now competes with a Treasury deficit near $2 trillion a year for the same private wallets, with the Federal Reserve no longer absorbing supply (Fortune, Jul 31).
Who profits in the meantime? The arrangers and the asset managers, visibly. Morgan Stanley collected $2.3 billion in debt and equity capital markets fees in the first half of the year, up from $1.4 billion a year earlier (LSEG data via press report, Aug 2026), and the six firms named in Nvidia's platforms stand between savers and the buildout, collecting management fees on assets valued by model rather than market. The hyperscalers profit too, in a narrower sense: every dollar of borrowing moved off the balance sheet is a dollar the rating agencies do not score. Oracle shows where the visible edge of this ends, having acknowledged in its June annual report that it cannot guarantee it will be able to manage its outstanding debt (New York Times, Jul 31).
The observable sequence if this read is right: spreads on AI-heavy investment-grade issuers drift wider relative to their euro-denominated twins, a divergence Karoui already flags as demand fatigue in the dollar market (CNBC, Aug 14), followed by the first failed or repriced GPU-backed vehicle refinancing well before the 2049 maturities. What breaks the read: hyperscaler revenue growth outrunning the obligations, so the take-or-pay payers keep paying from earnings rather than new issuance, and the hidden $1.65 trillion simply converts to ordinary boring debt without incident. Both outcomes remain live. Only one of them requires anyone to find out what the chips were worth.
The judgment the numbers earn is this: the bond market is being watched so closely precisely because it is the part of this boom that can be marked. Debt that appears in footnotes gets argued about at dinner parties; debt that lives inside insurance vehicles and uncommenced leases gets discovered in courtrooms. The visible half of the AI debt scares fund managers because they can sell it. The hidden half should scare them more, because by the time it has a price, someone's annuity already owns it.
The visible half scares fund managers because they can sell it; the hidden half should scare them more, because by the time it has a price, someone's annuity already owns it.
Method. This analysis rests on the sources cited below. ARCANE does not publish a proprietary universe, cohort weighting or exclusion list for this piece — the reading is the desk's, argued from the record, not a screened back-test.