The $370 Billion Put Nobody Is Being Paid For

Private credit has become the marginal lender to the AI buildout — and the entire structure rests on one instrument that the accountants say is worth zero, the rating agencies say is debt, and the bond market has already started charging 65 basis points more for.

1. Today's tape is the evidence, not the decoration

Monday 24 August, midday. The market is doing something very specific, and almost nobody is naming it.

XLF +1.27% against QQQ -0.75% is a 2.02 percentage point single-day spread on a day the index barely moved. Every entity that writes a guarantee in the AI credit chain is red. Every entity that collects a fee on the guarantee chain is green. The BDCs that actually hold middle-market loans are up half a percent while high yield sits perfectly flat — meaning this is not a generalised credit scare. It is a very precise repricing of who is short the option.

2. The primary source nobody read: BofA's own model of the guarantee

On 11 August, Bank of America's credit desk cut Broadcom's issuer and bond ratings to Marketweight from Overweight — explicitly on XPV platform concerns. The note is the most detailed public model of a residual value guarantee that exists, and it has been almost entirely ignored because it was published in a credit product, not an equity product.

Read what BofA actually assumed:

• Funding modelled ratably over 16 monthly rounds
• 5-year debt amortisation
• 20% price decline over the 5-year chip life
• 25% price shock at default at each point over the lease life

Output: maximum RVG exposure on the first tranche peaks at $26bn in September 2027, with maximum loss exposure of $2.9bn. Extended across the 20GW platform with 2GW deals added quarterly, maximum RVG reaches $370bn in mid-2029, with maximum loss of $42bn at a 100% default rate and $10.5bn at 25%.

Substitute a 50% five-year price decline for BofA's 20% and the loss figures do not rise by 30% — they rise by a multiple, because the guarantee is a first-loss position on the gap between resale value and outstanding principal. That is the whole point of a put.

What Broadcom itself says

Broadcom's own estimate, per the same note, is that the senior debt on the initial $35bn tranche carries maximum loss exposure of $29bn. Not $2.9bn — $29bn. That is the company's disclosure of gross exposure; BofA's $2.9bn is the modelled expected loss after their depreciation and default assumptions. The distance between those two numbers is the assumption set. Everything depends on it.

The 9 June 2026 launch release — the primary document — is worth reading for what it does not contain. Three CEO quotes across Broadcom, Apollo and Blackstone. Hock Tan on a "historic inflection point." Jim Zelter on "the capital foundation." Jon Gray on "financing through our credit and insurance business." The words "guarantee," "residual value," "backstop" and "contingent" appear nowhere in the release. The single structural feature that makes the transaction possible is absent from the announcement of the transaction.

3. The number I derived that nobody printed

Here is the calculation the sell side has not run.

The cross-check that makes me confident

There is a second, entirely independent measurement of the same thing. A Financial Times investigation, with Jefferies analyst Jonathan Petersen doing the arithmetic, found that data centre projects carrying a Google lease guarantee borrow at a median 7.1%, while neocloud operators building around Nvidia chips without such a backstop pay 9.3%. That is a 220 basis point credit-substitution wedge.

220bp from the FT/Jefferies dataset. 275bp from my Big Sky tranche arithmetic. Two completely different transactions, two completely different methods, both landing at roughly a quarter of a percentage point per hundred basis points of guarantee. That is not a coincidence — it is the market's price for converting an unrated AI lab into an investment-grade obligor, and it is remarkably stable at 220–275bp.

4. The number that says the market has already started charging more

Meta ran the same structure twice, and the second time cost more.

Same sponsor. Same guarantee. Same joint-venture split. A smaller and structurally simpler project. And the market charged 65 basis points more, nine months later. That is a 29% increase in the risk premium on an identical instrument, with nothing about the instrument having changed.

This is the single highest-information event on the calendar for the entire AI complex, and I said so on 30 July when I flagged "the spread on the NEXT hyperscaler SPV bond" as the tell to watch. I am now putting a number on it. If Broadcom's tranche 2 prints inside +280bp, the guarantee still works and my caution is wrong. If it prints wider than +320bp, the structure premium is decaying on schedule and every subsequent deal gets harder.

5. Where the paper actually ends up — and why this is a private credit story, not a tech story

Morgan Stanley's bridge is now the industry's shared arithmetic: roughly $2.9 trillion of global data centre capex through 2028, of which hyperscaler operating cash flow covers about $1.4 trillion, leaving a $1.5 trillion gap. Their allocation of that gap: ~$800bn from private credit, ~$200bn from corporate bonds, ~$150bn from securitised products.

Private credit is not a participant in the AI buildout. Private credit is the majority funder of the AI buildout.

And the paper does not stay with the funds. Sascha Steffen's 14 August analysis of Nvidia's $500bn platform traces the flow precisely, and it is the best single piece of work published on this subject:

"Follow the paper to its resting place and the answer is: insurance policyholders, several layers removed from the underwriting decision."

Of the six managers Nvidia signed non-binding MOUs with on 10 August — Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR — three own or control the insurers most likely to buy the paper. Apollo owns Athene. KKR owns Global Atlantic. Brookfield has Wealth Solutions. Blackstone's Credit & Insurance business anchored both the Big Sky tranche and CoreWeave's $8.5bn A3-rated delayed-draw term loan. PIMCO reportedly took ~$18bn of Meta's Hyperion.

6. The date nobody has circled: 30 January 2027

Here is the piece of this that I think is genuinely unpriced.

Insurance capital charges under both the US risk-based capital regime and Solvency II are keyed to ratings, not to collateral. And the European regime is about to get dramatically friendlier to exactly this paper, at exactly the wrong moment.

• Under Commission Delegated Regulation (EU) 2026/2692, applying from 30 January 2027, the revised Solvency II framework cuts the spread-risk capital charge on securitisations — illustratively, per KKR's own analysis, from about 46% to about 22% for a three-year non-STS position.
• Euro-area insurers already hold roughly €211bn of private credit, about 2.3% of assets, concentrated in Germany, France and the Netherlands and in a small number of large institutions (ECB Financial Stability Review, May 2026).
• Supervisory reporting cannot yet distinguish investment-grade asset-based finance from sub-investment-grade direct lending. A growing exposure to GPU-backed structures will show up in the data as generic high-grade fixed income.

Meanwhile the US is going the other way. The NAIC adopted new RBC factors for CLOs on 23 June 2026, effective with year-end 2026 reporting, and replaced the flat 6.8% collateral-loan charge with a framework keyed to what actually backs the loan, effective year-end 2027 — alongside a 45% charge on CLO residual tranches and closer scrutiny of offshore asset-intensive reinsurance.

7. The mechanism most people are missing: this is a rating event, not a default event

Every bearish take on AI credit I have read runs through default. Anthropic misses a lease payment; a neocloud folds; the guarantee gets called. That is the wrong model, and it is why the bears keep being early and wrong.

Steffen names the right one in a single sentence: "The stress scenario for this asset class is credit risk migration."

Work it through. Insurance capital charges are keyed to ratings. Fitch currently has an open consultation on how to treat residual values in these structures. If a rating agency changes its residual-value methodology — not because anyone defaulted, but because the committee revises an assumption about what a three-year-old TPU is worth — then:

1. The senior tranches get downgraded, or put on watch.
2. Insurance capital charges rise mechanically, with no discretion involved.
3. Rating-constrained holders become forced sellers.
4. They are forced sellers of an asset class with essentially no secondary market. Only ~$20bn of credit fund exposures traded in all of 2025 — roughly 1% turnover of the direct lending market. The credit secondaries bid can price a trickle. It cannot absorb a wave.

8. The 2029 convergence — three walls, one year, one page

This is the part I have not seen assembled anywhere, and it is the reason I am writing this piece today rather than next month.

Read those four rows together.

And the extensions are getting harder, not easier

The most under-read statistic in credit right now: of 2026 loan amendments, 30% went to issuers rated BB- or higher, up from 11% in 2025, while the B- share collapsed to 27% from 44%. Record amendment volume and a collapsing weak-credit share together mean one thing — the weakest cohort is being cut out of the extension market, not served by it.

Lincoln International estimates 30–40% of direct lending deals maturing in the next two years have already extended once, with leverage up roughly half a turn from inception across all vintages and closer to a full turn on surviving 2019–20 vintages.

9. What the underlying credit actually looks like right now

Strip away the AI story and look at the loan book that private credit already owns. Fitch recorded a US Private Credit Default Rate of 6.1% for the trailing twelve months ended July 2026 — a record high, up from 6.0% in May and April, with healthcare providers again the largest source of unique defaulters.

The default-rate debate is genuinely confusing, so here is the whole picture on one page:

Fitch's own composition explains the gap: of 32 private credit default events in Q2 2026 across roughly 1,300 tracked borrowers, more than half were maturity extensions rather than missed payments. A lender quoting a falling default rate is quoting hard credit events. A rating agency quoting a record is counting every accommodation a creditor was forced to grant.

Both are true. The difference between them is the entire subject. And the accommodations are running hot:

• PIK is on 10.6% of all direct lending loans, representing 8.9% of total interest income. Loans that carried no PIK at close but carry it today are 5.9% of all loans, against 2.5% at end-2021 — and that cohort's average LTV has gone from 39.4% at inception to 76.1% (Lincoln International).
• Of roughly 5,000 companies held by BDCs at 31 March 2026, 538 (10.6%) showed credit pressure, up 15% in a single quarter, with first-lien and unitranche investments under pressure up 44% to $35.4bn. Half of those 538 used no PIK in the preceding twelve months — meaning PIK-based screens miss half the problem.
• Direct lenders foreclosed on $24.2bn of principal in 2025 and a further $15.2bn in Q1 2026 alone, against $13.6bn across the preceding three years combined. One quarter at $15.2bn annualises to roughly two and a half times the whole of 2025.
• Commercial Chapter 11 filings in H1 2026: 4,589, up 28%. Subchapter V: 1,663, up 50%. In the same period the leveraged loan payment default rate fell.
• KBRA's tell: median interest coverage held at 1.6x, but the share of borrowers with improving coverage plateaued after two years of gains, and median EBITDA growth fell to 24% from 27% — the largest quarter-on-quarter decline in the series. Earnings growth has been doing the deleveraging work since 2023. That engine just started slowing.

And the funds themselves are still gated

Non-traded BDC redemption gates were still binding in Q2 2026. Two large vehicles fielded $4.7bn of tender requests in the quarter. Where requests reached 18.8% and 38.1% of shares outstanding against a 5% quarterly cap, roughly 27% and 13% of each shareholder's request was honoured.

10. Where I disagree with consensus, stated plainly

Consensus view #1: "The guarantees are nearly costless."

Wrong, and demonstrably so. Bloomberg framed the backstop as "something akin to a free lunch." CreditSights got closer: Nvidia is "writing a put... nearly costless in the boom phase, but becomes most relevant in a severe, abrupt downturn." Even CreditSights understates it. The guarantee is not costless in the boom phase. It costs 275 basis points a year in foregone price, which is roughly 3.3% of the revenue it buys and 4.8% of the gross profit. It is being paid for continuously, in a currency (spread compression handed to the customer) that no accounting standard requires anyone to disclose.

Consensus view #2: "Nvidia's financing platform is evidence of demand strength."

I read it as the opposite. Jefferies documented that Google-guaranteed projects borrow 220bp cheaper than Nvidia-ecosystem projects — roughly $770m a year on a $35bn facility. Nvidia's response was the $500bn platform, up to $125bn of residual support, and $105bn behind a single Ohio campus. That is not a demand-generation move. It is a defensive refinancing of a 220 basis point cost-of-capital disadvantage that Google had already built. Nvidia's customers were being outbid on the price of money, and Nvidia closed the gap with its own balance sheet. That is a materially different thing from strength, and it is why Nvidia's 5-year CDS has behaved the way it has.

Consensus view #3: "Buy the crypto miners turned AI landlords — they got cheap Google-guaranteed debt."

No. Google already took the value. TeraWulf used a Google lease guarantee to turn a 360MW project into a $3.2bn construction bond — and handed Google penny warrants for roughly 14% of its equity. Cipher Digital got $1.4bn of lease obligations backstopped on a 168MW Texas facility and gave up warrants for roughly 5.4%. Hut 8's $7bn River Bend deal reportedly came without warrants, which makes it the interesting one.

The cost-of-capital advantage is capitalised into the guarantor's warrants, not into the host's equity. If you are buying a Google-backed host because it borrows at 7.1% instead of 9.3%, understand that you are buying an entity that sold 14% of itself for that privilege. Do the arithmetic before you assume the trade.

Consensus view #4: "AI credit risk is a Nasdaq problem."

It is an investment-grade bond index problem. The six largest tech names already account for roughly 8.6% of duration-times-spread risk in the US IG corporate index. Layer on top of that the SPV paper sitting in insurance general accounts and the securitisation charge halving in Europe on 30 January 2027, and the honest statement is this: anyone holding a total bond market fund is long the AI buildout, and did not choose to be.

11. Three scenarios, with measurable triggers

Bull — 25%

Trigger: Broadcom's tranche 2 senior slice prices inside +280bp with oversubscription above 2x; Fitch's residual-value consultation lands crediting front-loaded amortisation within contract tenor; Solvency II recalibration applies on schedule 30 January 2027; 30Y closes below 5.00%.
Consequence: the funding channel widens rather than narrows. AVGO and NVDA recover their guarantee discount. The managers re-rate harder than the guarantors — APO to $165–175, BX toward $175. BDC discounts close from ~21% to ~12%; OBDC reaches $13.20 inside three months. This is the scenario where I am wrong about the direction and still right about the pair.

Base — 55%

Trigger: tranche 2 prices +280 to +330bp, gets done, but with a fatter junior tranche and a larger Broadcom backstop as the price of getting it done. S&P continues adding RVGs to adjusted debt. More sell-side credit desks follow BofA to Marketweight. No default anywhere.
Consequence: the guarantors' equity multiple compresses while their earnings keep growing — AVGO at 68x trailing and ~87x forward is the vulnerable number, not the revenue line. The managers' fee streams compound. BDC discounts persist without widening. The pair trades work; the outright directional bets do not. This is the scenario I am positioned for.

Bear — 20%

Trigger — any one of four, all measurable: (a) Anthropic's October IPO prices below the $965bn post-money mark set by its Series H; (b) any neocloud or AI lab misses a scheduled lease payment; (c) Fitch publishes a residual-value methodology change that downgrades or watch-lists existing SPV senior tranches; (d) 30Y UST closes above 5.50%.
Consequence: the RVG stops functioning as a rating enhancement. Insurance holders become rating-constrained forced sellers into an asset class with ~1% annual turnover. The contagion runs into IG bond funds, not out of the Nasdaq. AVGO to $270 (roughly 15x a haircut earnings estimate), OBDC to $9.50, OWL to $8.00, and the credit secondaries bid — which cleared BlackRock TCP's book at 95 cents in August — gaps to the low 80s.

12. The trades

13. Prior-call scorecard — including what is not working

14. Cross-asset regime levels

15. What to do this week

1. Do not size anything fully before Wednesday. NVDA reports 26 August, MRVL 27 August, and Warsh gives his Jackson Hole debut on 28 August. Three sequential binaries in three days.
2. Put [RVG-1] on at half weight today at a 2.72x ratio. Add the other half only after Broadcom's early-September print, or on a bounce to 2.85x.
3. Buy APO in $126–136. It is in the zone now. This is the position I want largest.
4. Hold OBDC. Add only below $11.00. Add the 3%-of-fair-value AI-equipment-finance tripwire to your monitoring, not just the non-accrual line.
5. Diarise 30 January 2027. The Solvency II securitisation charge cut from ~46% to ~22% is the single largest scheduled change in demand for this paper and it is not in anyone's model.
6. Set a standing alert on Fitch's residual-value consultation. A methodology change is the bear trigger that requires nobody to default.
7. Read Broadcom's tranche 2 pricing as the answer, not the noise. Inside +280bp: the guarantee works. Wider than +340bp: flatten.

The thesis in one paragraph

Private credit has become the majority funder of the AI buildout — roughly $800bn of a $1.5tn financing gap — and it is doing so against collateral that depreciates faster than the debt amortises. The instrument that makes this possible is the residual value guarantee, and its market price is observable at 275 basis points from the wrapped-versus-unwrapped tranches of the Big Sky deal, cross-checked at 220 basis points by the Jefferies Google-versus-Nvidia funding wedge. That price is a real, continuous, annual cost to the guarantor — roughly 3.3% of the revenue it buys — and it appears nowhere in reported gross margin, which is why the equity market values Broadcom as though the guarantee were free while the bond market has already docked it 30–45 basis points and Bank of America has cut it to Marketweight. The bond market resolves this class of disagreement first. Meanwhile the guarantee itself is repricing at 7.2 basis points a month — Meta paid 65bp more for the identical structure nine months later — and Broadcom's maximum exposure peaks at $370bn in mid-2029, the same year the rescheduled leveraged loan wall and the BDC maturity wall land, by which point the collateral will be two silicon generations old. The right trade is not to short the AI buildout. It is to own the entities that collect fees on the guarantee chain and to fade the entities that wrote the guarantee, because the accountants say the put is worth zero, the rating agencies say it is debt, and only one of them will turn out to be right.