Broadcom (AVGO) Falls 5% on a $370 Billion AI Debt Estimate: The $29 Billion in the 10-Q Is the Real Exposure
Broadcom traded down more than 5% Friday, touching roughly $390 after opening above $411, shedding about $102 billion of market value across 4.76 billion shares. The trigger was not a guidance cut, a customer loss, or a downgrade. It was an estimate. Bank of America’s Tom Curcuruto calculated that the chip-financing vehicle standing behind Broadcom’s AI expansion could carry $370 billion of senior debt by mid-2029 at 20-gigawatt scale, including roughly $150 billion of new issuance in 2027 alone, assuming the platform grows at about two gigawatts per quarter. The number is enormous, it is not Broadcom’s debt, and the market sold the stock as though it were.
The distinction matters more than the headline. The AI XPV Platform took shape in June, when Apollo and Blackstone anchored a $35 billion institutional financing designed to fund more than one gigawatt of compute for Anthropic, with the structure sized to support more than twenty gigawatts for frontier labs through 2028. Blackstone has since sounded out investors on a follow-on exceeding $30 billion. The vehicle, not Broadcom, raises the debt. Institutional capital buys the racks and leases them to the AI customer, which solves the problem that has actually been throttling deployment: demand for custom accelerators far exceeds the capacity of the labs to fund the upfront cost from their own balance sheets. Broadcom booked over $30 billion of AI orders last quarter against $10.8 billion shipped. The financing platform is the mechanism that turns the first number into the second.
What Broadcom does own is the backstop. The latest 10-Q discloses that an investor partner assumed the rack purchase agreements and the associated customer leases, and that Broadcom agreed to backstop those lease payments for five years, with maximum exposure of up to $29 billion on the initial transaction. That figure assumes a complete customer default with zero recovery value on the equipment, which is not a scenario anyone underwrites. If a customer does default, Broadcom can assume the lease itself or arrange a sale of the hardware, and either action reduces what it ultimately pays. Run BofA’s own math to its conclusion and a full default across a twenty-gigawatt platform produces losses near $42 billion against $370 billion of notional. The guarantee is roughly eleven cents on the notional dollar. Sell the $370 billion headline and you are selling the wrong number by a factor of nine.
The number worth arguing about is the coverage ratio, and it is the reason the market split its verdict on AI financing inside a single session. Nvidia’s $500 billion framework carries an option to backstop up to $125 billion, or twenty-five percent. Broadcom’s disclosed maximum sits at $29 billion against an initial $35 billion transaction. The bases are not identical — one is a discretionary participation ceiling, the other a theoretical worst case on a signed deal — but the shape of the difference is real. Broadcom is standing much closer to the paper than Nvidia is. That is what a challenger does when it needs to convert a design win into a deployment, and it is what a market leader does not have to do. AMD rose 4% to roughly $503 the same afternoon on a Street-high $1,250 target from Baird; Nvidia finished about flat; Marvell fell 1.3%. Vendor-backstopped financing read as a growth enabler for one name and a liability for another on the same day, on the same theme.
The residual-value guarantee is where the analysis gets uncomfortable, and it has nothing to do with credit quality. Broadcom is short a put on the second-hand value of its own accelerators, with roughly $30 billion of initial residual-value guarantees and credit support attached to the senior notes. The counterparty risk that dominates the discussion — will Anthropic pay its lease — is the smaller half. The larger half is that Broadcom’s own product roadmap is the primary agent of impairment. Every generational XPU it ships lowers the resale value of the racks it has guaranteed. A company that has spent thirty years selling upgrade cycles has now written a five-year commitment that the previous cycle retains its worth. The bull rebuttal is empirical and reasonably strong: Nvidia’s six-year-old Ampere fleet remains fully utilized with resilient token pricing, and occupancy on legacy hardware has held even as new silicon arrived. Compute has not yet behaved like a depreciating asset. It has behaved like a scarce one. The guarantee is a bet that inference demand keeps absorbing whatever the frontier discards, and that bet is currently winning.
The moat argument cuts in the same direction and gets underweighted. Broadcom’s defensible position is not the XPU design business, which a hyperscaler can in-source, but the networking attach that survives that decision, and now the balance sheet itself. Adjusted EBITDA margins guided near 68%, $10.26 billion of free cash flow in the quarter and $19.63 billion of cash on hand make the backstop a product feature. Marvell cannot offer it. A merchant ASIC startup cannot offer it. When a frontier lab chooses between silicon it can buy and silicon it can lease against a vendor guarantee, the second option wins on cash flow timing alone. Financing capacity has become part of the competitive stack, and only two companies in the sector have enough of it to matter.
The stock has been the odd one out for a quarter. Broadcom closed at $495 the day of its Q2 print on June 3, beat cleanly, and drifted to the low $400s. It sits up about 21% year to date and 36% over twelve months at roughly a $2.0 trillion capitalization, against AMD’s 126% year-to-date advance and Marvell’s 26% three-month gain. Forward P/E near 69 is the constraint every note keeps returning to. The sell side has not followed the tape down: consensus is $527.88 across 48 analysts with 44 Buy ratings and no Sells, the range runs from $215.88 to the $675 that BNP Paribas Exane raised to on Friday, and Goldman’s removal from its Conviction List on August 3 and Erste’s cut to Hold are the only visible defections. That gap between a $527 consensus and a $390 tape is not a forecasting disagreement about the quarter. Nobody doubts the quarter. Q3 is guided to $29.4 billion, up 84% year over year, with AI semiconductor revenue guided to $16 billion against $10.8 billion delivered, and prediction markets assign a 94% probability to clearing $15 billion. The gap is a disagreement about what multiple belongs on revenue that the seller helped finance.
Base case is $380 to $440 into the September 2 print, with the AI number beating and the multiple refusing to expand because every incremental dollar of guided revenue now arrives with a guarantee attached. Bull case runs to $500 on a Q4 guide above $20 billion of AI semiconductor revenue plus explicit disclosure that platform economics improve as scale reduces the required backstop percentage — the argument that the $29 billion coverage ratio was a first-deal premium and not a structural term. Bear case is cohort derating to $300 to $330, and it does not require a customer default. It requires only that the market decide vendor-financed AI revenue deserves the multiple of a leasing business rather than a semiconductor franchise, at which point 69x forward compresses toward the low 40s regardless of what the accelerators do. Fiscal 2026 AI semiconductor revenue guided to $56 billion and fiscal 2027 guided above $100 billion do not defend against that; they enlarge the base the derating applies to.
September 2 will produce a revenue number everyone already knows and a disclosure almost nobody is watching. The line to read is the maximum exposure figure on lease and residual-value guarantees in the quarterly filing. If it stays near $29 billion while the platform adds gigawatts, the first transaction was priced as a launch subsidy and the structure scales without Broadcom absorbing more of it. If it climbs proportionally with committed capacity, then BofA’s $370 billion is the right denominator after all and the equity is being repriced for the correct reason.