The $200 Billion of AI Debt Hiding Inside Your Income Fund

Nvidia's equity just printed an all-time high. Its credit counterparties are marking down. Blue Owl's technology lending fund has lost 6.8% of NAV in seven months while paying a 9.28% distribution — a negative total return, in first-lien senior secured paper, during the greatest technology boom in history. That is not a coincidence. It is the tell.

1. The One-Sentence Thesis

2. What Actually Happened Today

Nvidia reported Q2 FY2027 last night and the stock is up 8.28% to $227.02. Look at what moved with it, and what did not:

Three things fall out of that table, and each one carries a consequence.

First: the alternative asset managers are no longer financials. OWL, ARES, BX and APO all beat XLF by 120 to 350 basis points on a day when the catalyst was a semiconductor earnings report. Blue Owl outperformed the financials sector by 348bp on news that had nothing to do with lending spreads, deposit costs, or credit. The market has quietly reclassified these firms as AI-infrastructure beta. That matters because it means their multiples now carry AI-cycle risk premium that their fee streams do not actually earn — and when the AI credit cycle turns, they will de-rate on both the numerator and the denominator.

Second: the borrowers did not participate. Nvidia rose 8.28%. CoreWeave — which exists to buy Nvidia GPUs with borrowed money — fell 0.51%, an 879bp single-session divergence, and closed near its session low after opening at $91.34. Oracle did the same thing. When the chip supplier's blowout quarter is a non-event for the entities that owe money for the chips, the credit market is telling you the marginal dollar of GPU demand is now being financed, not funded. That is a distinction with a maturity date attached.

Third: this happened on a hawkish day. Cleveland Fed President Beth Hammack said from Jackson Hole this morning that "now is the time to act" on raising interest rates, with Chair Kevin Warsh's keynote tomorrow. Private credit is a floating-rate asset class. Rate hikes are not a headwind to lender income — they are a solvency event for the borrower. I quantify that in Section 6.

3. The Primary Source: Nvidia's Own 10-Q Is a Credit Document

The headline was a beat-and-raise. The filing was something else. From CNBC's read of the quarterly filing and sell-side notes published this morning:

Goldman Sachs' James Schneider tallied management's total financial commitments in support of customers at $366 billion: $279bn supply/capacity (largely memory), $29bn cloud service agreements, $25bn data centre leases, $25bn equity investments, $5bn capex. Truist's William Stein attributed the DSO extension directly to Nvidia "offering extended payment terms to investment grade customers."

Read that as a lender would. Free cash flow halved not because the business slowed but because $24.6 billion of revenue moved from cash into receivables in six months, concentrated in five counterparties. Nvidia is not just selling GPUs. It is financing them — through payment terms, through the $500bn financing platform announced this month with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR, and through the July 1 AI Compute Partnership that backstops neocloud GPU deployments and commits to rent back unused capacity at a fixed rate.

Consequence one: a vendor that guarantees the residual value of its own product has converted a sales risk into a credit risk, and it has not moved the risk off the system — it has concentrated it. Consequence two: the counterparties who took the other side of that guarantee are private credit funds, and their investors are retail.

4. Where the Paper Actually Sits

The lender base for the AI buildout is not the banking system. Per the Quinn Emanuel private credit client alert, outstanding private credit loans to AI-related companies have gone from near zero to over $200 billion, with Morgan Stanley projecting a further $800 billion over the next two years. The Chicago Fed's February 2026 study found direct bank exposure to AI-adjacent industries averaging just 0.8% of total assets — while explicitly noting that additional exposure runs through channels not captured directly. Those channels are the funds.

GQG's work on the buildout puts the total data centre financing gap near $1.5 trillion, with roughly half expected to come from private credit. Independent analysis of the data centre debt stack sizes the current market around $500bn and documents its dominant structures: SPVs, delayed-draw term loans, and variable interest entities — i.e. off-balance-sheet, project-financed, and collateralised by hardware.

I went and found it in a filing

This is not abstract. I pulled Blue Owl Capital Corporation's 10-Q for the quarter ended June 30, 2026 and read the Schedule of Investments. Inside a publicly listed, retail-owned BDC yielding 11.6%, alongside the middle-market manufacturers and healthcare roll-ups you would expect, sit these positions:

• Nscale Services UK Ltd — first lien senior secured delayed draw term loan
• Nscale Global Holdings Limited — preferred equity and Series B preferred shares
• Space Exploration Technologies Corp. — Class A common stock

Nscale is a European AI/GPU cloud operator. Per Fortune's June profile it has raised roughly $4bn including a ~$1.4bn GPU-collateralised loan and a Norwegian facility, and is burning cash against contracts with the largest AI labs. Separately, Blue Owl portfolio company Wingspire Equipment Finance closed a $140m large-equipment financing for a private-equity-backed provider of GPU cloud capacity to AI labs.

Three structural problems in one holding. (1) It is a delayed draw — an undrawn contractual commitment to lend more to a borrower that may be deteriorating. That is precisely the line item that sank Nortel in 2001, where the funded balance halved in a year not through repayment but through write-down, while the unfunded commitments could not be withdrawn. (2) The BDC holds both the senior debt and the preferred equity of the same borrower — the lender is also the sponsor, which is exactly the conflict that makes a mark unreliable when it matters. (3) The collateral is GPUs, whose useful life is a disclosed accounting assumption, not a physical fact.

The collateral disagreement

On comparable hardware, CoreWeave depreciates over six years; Nebius, running a comparable business, uses four. That is a 50% spread in annual expense against the same asset class — and the operator with the longest assumed life is the one whose hardware is pledged. In early 2025, Meta extended its useful lives while Amazon shortened a subset of its fleet, citing faster technology turnover. Depreciation cannot move in two directions at once if it is measuring something physical. Somebody is wrong, and whoever is wrong has been booking the difference as earnings — and, more importantly here, as loan-to-value.

CoreWeave carried $35 billion of total debt at June 30, 2026, with quarterly net interest expense of $640 million versus $267 million a year earlier — a 140% increase — against a $104 billion backlog. Much of the sector's debt matures between 2026 and 2028. Oracle's five-year CDS has traded near 200bp against roughly 78bp for Nvidia and 93bp for Meta. The credit market has already separated the guarantor from the guaranteed by 122 basis points.

5. The Finding Nobody Is Talking About

Here is where I depart from consensus. The AI-lending exposure above is prospective risk — real, growing, and mostly not yet impaired. But there is a second AI credit exposure that is impairing today, and it is showing up in monthly NAV disclosures that anybody can read and almost nobody does.

Blue Owl runs two large non-traded BDCs with near-identical fee structures, vintages and distribution rates. One lends to technology companies. One lends across the economy. Compare them.

Source: monthly NAV histories published at otic.com and ocic.com. Total return estimated as NAV change plus 7/12 of the stated distribution rate.

Read the bottom four rows carefully, because this is the entire argument.

In 2025, the technology book was the better book — OTIC's NAV was essentially flat (−0.48%) while the diversified fund lost 2.31%. Lending to software was the highest-quality thing you could do in private credit, and it had been for a decade: recurring revenue, high gross margins, sticky customers, low capital intensity.

In 2026 the sign flipped. OTIC is down 6.84% in seven months against OCIC's 2.68%. That is a ~600 basis point swing in relative NAV performance in twelve months, between two funds run by the same manager, with the same underwriting committee, the same leverage policy and the same 9.2-something distribution rate. There is no fee explanation, no rate explanation and no vintage explanation. The only variable that differs is what the borrowers do for a living.

And note the single worst month: January to February 2026, $10.27 → $9.97, −2.92% in thirty days. On a book that is 87% first-lien senior secured, at typical non-traded BDC leverage of roughly 1.0× debt-to-equity, a 2.9% equity NAV decline implies about 1.4–1.5% of gross assets written down in a single month — the equivalent of marking roughly 5% of the loan book down 30 points, at once. Rates did not move 1.5% of a first-lien book in February. That was credit.

6. The Rate Mechanism — Quantified

This morning, Hammack argued for hikes. Separately today, Benefit Street Partners' global investment strategist Anant Kumar warned that private credit lenders must underwrite "a wider distribution of interest rate outcomes than the forward curve implies," and treat rate volatility, not just the level, as a risk, noting that with forward guidance scaled back, "sell-offs become self-fulfilling spirals without reassuring Fed commentary to lean against." He also flagged the fiscal backdrop: 6% deficits into a warm economy, with the long end "demanding fiscal correction" and receiving "jawboning and a $4bn buyback."

Here is the arithmetic that makes this a solvency question rather than an income question.

A typical direct lending borrower carries 6.5× debt/EBITDA. Debt equal to 6.5× EBITDA means every 100bp of rate increase consumes exactly 6.5% of EBITDA in incremental cash interest. At 7.5× leverage, 7.5% of EBITDA per 100bp.

This is Minsky's mechanism, and the point is that nobody has to behave badly for it to happen. A borrower moves from "speculative" — can service interest, must refinance principal — to "Ponzi" — cannot service interest from operations — without changing a single thing about how it runs. The rate moves, or the asset gets remarked, and the classification changes underneath them. Two things are currently in motion at once: the cost of refinancing and the carrying value of the collateral. That combination, not a collapse in AI demand, is what breaks this.

Note what the historical record says here. The bear case does not require AI to fail. Insull's power plants never missed a day; the holding companies above them died anyway in 1932 when refinancing was simply refused. Nortel's carriers wanted the equipment; the vendor financing that made the sales possible reversed the receivable, the revenue, and the goodwill simultaneously. In every case the failure was at the settlement date, not the demand curve. That framing is developed well in this Unicus Research piece from August 17, which is the best thing I have read on the structural precedent.

7. Sector Scoreboard

Against that backdrop, here is where the industry's own credit metrics stand.

The Financial Stability Board issued a formal warning on private credit vulnerabilities in May 2026, and Fitch flagged BDCs as under pressure in Q1. Neither is priced anywhere in the manager equities that rallied today.

The arbitrage that should not exist

Now put the listed and non-traded vehicles side by side. Same asset class. Same seniority. Nearly the same yield.

Blue Owl's own listed BDC trades at 0.796× book while Blue Owl's own non-traded BDCs transact at 1.000× book. Same manager. Same origination platform. Same credit committee. A ~20% valuation gap on economically similar paper, sustained only because one vehicle has a public price and the others have an appraisal.

There are exactly two ways that gap closes: listed prices rise, or non-traded NAVs fall. There is no third option. Retail money is overwhelmingly parked on the 1.000× side.

8. Positioning — My Recommendations

9. Three Scenarios

10. What Would Prove Me Wrong

I will hold myself to these, explicitly:

• If OTIC's NAV stabilises and rises for two consecutive months, my "AI is eating the software loan book" thesis is wrong and I will say so. That is the load-bearing observation in this note.
• If OBDC's Q3 non-accruals print above 2.5% at fair value, the 20% discount was information, not opportunity, and I was reading the marks too generously.
• If non-traded BDC NAVs hold flat while listed discounts narrow to single digits, then the appraisal marks were right all along, the listed market was simply mispricing liquidity, and my pair trade earns its return for the wrong reason.
• If Nvidia's next 10-Q discloses a contingent backstop obligation that is small — rather than the $366bn of gross commitments implied today — the circularity concern is materially overstated.

11. Bottom Line