Bloomberg reported on August 20 that Broadcom is in talks with lenders to raise more than $60 billion of debt for an AI chip financing arrangement benefiting Anthropic and other companies — a roughly $30 billion junior tranche, with Broadcom guaranteeing part of a senior-secured tranche of $60–70 billion, potentially taking the whole thing to $100 billion. Blackstone and Apollo are said to be in talks to participate. An X news module carried it as "Broadcom Secures Over $60 Billion Debt for AI Chip Deals" across 269 posts.
Two things about that paragraph. It describes negotiations, not a transaction — "in talks with lenders" is not a raise, and $60B/$70B/$100B is a range under discussion. And we have not read the Bloomberg story, which sits behind a paywall; every figure in it reaches us through secondary coverage and is treated as press-reported throughout, including the word "secures", which is doing work the reporting does not support.
Nothing below depends on any of it. Because Broadcom already filed the version of this story that can be checked, and it is a different structure with a smaller and much more specific number in it.
What the filing actually says
Broadcom's Form 10-Q for the quarter ended May 3, 2026 carries the arrangement twice — once in Note 11, Subsequent Events, and again verbatim in Item 5, Other Information. In full:
"On June 8, 2026, we arranged for Apollo ("investor partner") to take on certain agreements to purchase AI racks based on custom AI accelerators designed by us and the related lease agreements with a customer that enable access to compute capacity. In connection with the arrangement, we entered into a backstop agreement with the investor partner for the customer's lease obligations over 5-year terms. The backstop will increase over time as the AI racks are deployed and decrease as the customer makes payments on its lease obligations, with a maximum exposure of $29 billion. In the event of default by the customer, we have various remedies, including the assumption of the lease or effecting a sale of the AI racks, which would reduce our maximum exposure."
Read that against the headline and four things change.
It is a lease backstop, not a debt guarantee. Broadcom is not standing behind borrowings. It sold racks to Apollo, Apollo leases them to a customer, and Broadcom stands behind the customer's lease payments over five-year terms. That is a residual and payment backstop of the kind capital-goods vendors have written for a century — closer to a manufacturer supporting an equipment lease than to a chip vendor co-signing a loan.
The number is disclosed and it is capped. $29 billion, stated as a maximum, on a balance that rises as racks deploy and falls as the customer pays. It is not what Broadcom owes; it is the ceiling on what it could ever owe, before any recovery.
The filing does not name Anthropic. It says "a customer". It names Apollo and not Blackstone. It gives no size for the June arrangement, no vehicle, and no platform name. Every one of those attributions is press-reported, and the difference between "a customer" and a named borrower is exactly the distinction that decides whose obligation this is.
Broadcom told shareholders this was coming, in the risk factors of the same document. Customers "may make and have made greater demands on us with regards to pricing and contractual terms, such as seeking to lease our custom AI accelerators or XPUs or requiring us to purchase and then lease to them full AI racks or systems based on our XPUs instead of purchasing these chips or racks directly, as well as alternative financing arrangements for such leases or other novel or deferred payment models." And: "such arrangements have and may impose financial obligations, including backstops or guarantees, upon us or increase our exposure to credit or customer default risks."
That is the company describing, in its own filing, the business model change the wire story is reporting as news.
The denominator nobody attached to it
A $29 billion contingent exposure is either large or small depending on what it sits against, and the same 10-Q supplies all three denominators.
| Amount | Backstop as a share | |
|---|---|---|
| Total debt on the balance sheet | $64,907M | 44.7% |
| Net debt after $19,628M of cash | $45,279M | 64.0% |
| Book equity | $87,691M | 33.1% |
| Free cash flow, trailing four quarters | $32,762M | 0.89× |
The last line is the one worth holding. Broadcom's maximum exposure on this arrangement is about ten and a half months of its own free cash flow — 0.89× the $32,762M it generated across the four quarters to May 3. That is a real obligation and it is not a solvency question, which is a distinction the headline number does not let a reader make.
Against the reported figures the ratios invert, and this is the chart above in one line. The press-reported $60 billion is 0.92× Broadcom's entire balance-sheet debt and 1.83× its trailing free cash flow; the $100 billion top of the range is 1.54× and 3.05×. But those are the size of a financing vehicle, not of Broadcom's obligation, and treating them as Broadcom's is the same category error we took apart in the Marvell warrant piece yesterday, where a $12.18 billion aggregate exercise price was reported as a stake. The quantity that belongs on Broadcom's balance sheet discussion is $29 billion, it is a maximum rather than a balance, and it has been public since June 9.
The same week, the third structure
This is the fourth AI financing we have written about in three days, and the four are not variations on one thing. They put the risk in four different places.
| Structure | Size | Who carries it |
|---|---|---|
| Nebius funding stack | ~$21.4B | Customers, for 42% of it — prepayments the supplier gets in advance |
| CoreWeave debt and leases | $35.1B debt, $64.6B rent | The operator: $31.4B of the debt is recourse |
| Marvell–Google warrant | $12.18B strike notional | Nobody yet — it vests against $120B of revenue |
| Broadcom's Apollo backstop | $29B maximum | The chip vendor, contingently, capped, on someone else's lease |
The Nebius case is the inverse of Broadcom's: customers pay the supplier up front, funding 42% of the 2026 stack. Broadcom's is the supplier standing behind the customer instead. CoreWeave is the case where the obligation is real, recourse and simply larger than the debt note everyone reads. All three are vendor-adjacent finance, and only Broadcom's puts the vendor's balance sheet behind the customer's credit.
What separates Broadcom's from the other three is not the direction. It is that Broadcom's is the only one where the exposure is capped and stated in a filing, and the only one where the market is trading a number — $60B, $100B — that no filing contains.
What it does to our model, which is less than you would expect
We published a Broadcom model today. Its base case puts fair value at $282.09 a share against a price of $362.48 as of August 19 — a build-time snapshot, not a live quote — so the model already says the stock is about 22% expensive before any of this.
The backstop does not appear in it, and it should not appear as revenue. The sale to Apollo is a sale: the racks are shipped and the revenue is booked, which is part of why the guided ramp is deliverable at all. What the backstop is, in valuation terms, is a contingent claim on equity.
So bound it the hard way. Assume the customer defaults, assume the backstop is drawn to its stated maximum, and assume zero recovery — no lease assumed, no racks sold, none of the remedies the filing says would reduce the exposure. Charge the whole $29 billion against equity value across the 4,876 million diluted shares Broadcom reported:
| Charged against the base case | Per share | Share of the $282.09 base |
|---|---|---|
| $29B backstop, drawn in full, zero recovery | $5.95 | 2.1% |
| A second backstop the same size, also fully drawn | $11.89 | 4.2% |
| One turn off the exit multiple, 7× to 6× | $28.90 | 10.2% |
The worst case on the entire disclosed backstop is $5.95 a share. One turn of the terminal revenue multiple is $28.90 — 4.9× larger. Even a second arrangement of identical size, drawn in full with nothing recovered, moves the fair value less than half as much as the multiple assumption we chose ourselves.
Two things follow. The model does not change, and the size of the "no" is the finding: a credit story that costs 2% of fair value in its worst case is not what makes this stock cheap or expensive. What makes it cheap or expensive is whether AI revenue plateaus near the $137 billion a year our model settles on, and what a plateau is worth — and that $28.90 is our own assumption, not a disclosure, which is precisely why it deserves more scrutiny than the $29 billion does.
The honest caveat is that this bound only covers what has been filed. If a second vehicle produces a second backstop, the arithmetic above scales linearly and stays small; if it produces something structurally different — an on-balance-sheet guarantee of borrowings rather than a capped lease backstop — the bound does not hold and neither does this section. That is the disclosure to wait for, not the wire story.
What to watch
- Whether an 8-K or the Q3 10-Q gives a second backstop a number. The June arrangement was disclosed nine days after it was struck. If a $60–100 billion vehicle carries Broadcom credit support, the same disclosure obligation applies, and the figure to compare is a maximum exposure against the $29 billion already filed.
- Whether the backstop stays contingent or becomes recognised. The 10-Q carries it as a subsequent event with a maximum exposure and no liability booked. A guarantee that moves onto the balance sheet is a different accounting fact, and it would show up in net debt rather than in a note.
- Whether Broadcom names the customer. The filing says "a customer". Concentration in the AI segment is already the risk our model's notes call out and no slider expresses; a named counterparty for a five-year, $29 billion-capped lease obligation would let a reader size it.
- Anthropic's S-1. A confidential draft has been at the SEC since June 1. A company preparing audited public accounts that is also the reported beneficiary of vendor-supported purchase and lease obligations will have to disclose committed purchase obligations, and that will be the first time the other side of this arrangement is visible.
- What BofA's $370 billion figure actually measures. It is reported as the senior debt a scaled version of this platform could support by 2029 — a projection about a vehicle, not Broadcom's debt — and it moved the stock 5–6% when it circulated. The definition matters more than the number.
The June 8, 2026 arrangement — Apollo as investor partner, the purchase of AI racks based on Broadcom-designed custom AI accelerators, the related lease agreements with an unnamed customer, the backstop over five-year terms, the $29 billion maximum exposure, the increase-and-decrease mechanic and the default remedies — is quoted verbatim from Broadcom's Form 10-Q for the quarter ended May 3, 2026, where it appears in Note 11 and again in Item 5. The risk-factor language about customers seeking leases and alternative financing arrangements, and about backstops or guarantees imposing financial obligations, is from the same document. Total debt of $64,907M ($2,252M short-term and $62,655M long-term), cash of $19,628M, total assets of $179,158M and total liabilities of $91,467M are from the balance sheet in the Q2 FY2026 earnings release of June 3, 2026, as are the 4,876 million diluted shares and the quarterly free cash flow figures. Net debt of $45,279M, book equity of $87,691M, trailing free cash flow of $32,762M (the four quarters to May 3, 2026) and every ratio in this piece are R40 arithmetic on those filed figures. Broadcom's Q2 AI semiconductor revenue of $10.8B and the $16.0B Q3 guide are from our Q2 FY2026 analysis. The reported talks — more than $60 billion of debt, a roughly $30 billion junior tranche, a $60–70 billion senior-secured tranche, a potential $100 billion total, Blackstone's participation, Anthropic as beneficiary, a $35 billion June financing and BofA's $370 billion figure — are press-reported, describe negotiations rather than a completed transaction, and appear in no Broadcom filing we have read; we have not read the Bloomberg article itself, which is paywalled, and these figures reach us through secondary coverage and a trending news module. Anthropic's confidential draft S-1 submitted June 1, 2026 is press-reported. Fair values of $282.09 in the base case and $253.19 at a 6× exit multiple are computed from our Broadcom model as committed; the exit multiple, the terminal revenue path and the $137 billion terminal AI run rate are our assumptions and not disclosures. The $362.48 price is the build-time snapshot committed in this repo as of August 19 and will differ from the live quote on Broadcom's stock page. Nebius and CoreWeave figures are from our convertible-accretion piece and lease analysis; the Marvell warrant figures are from yesterday's piece.
