On August 11th, Tom Curcuruto, a credit analyst at Bank of America, issued a short downgrade report. He lowered his rating on Broadcom's bonds from "Overweight" to "Market Weight."
Note, this is not a stock rating, but a credit-side rating—meaning he is not worried about whether Broadcom is profitable, but whether Broadcom's balance sheet can withstand what it is doing.
There is a number in the report: $370 billion.
This is his estimate of the maximum Residual Value Guarantee (RVG) exposure Broadcom could assume by mid-2029 under a financing arrangement called the "AI XPV Platform." If all clients default simultaneously and all chips become worthless, Broadcom's theoretical maximum loss could reach $42 billion.
Within three days of the news, Broadcom's stock fell nearly 6%, breaking below $400, retreating more than 20% from its June all-time high.
What does $370 billion mean? Broadcom's entire market capitalization is about $2 trillion, with total on-balance-sheet debt of $64.9 billion.
A contingent, off-balance-sheet guarantee exposure could balloon to nearly one-fifth of the market cap!?
XPV
On June 9, 2026, Broadcom, Apollo, and Blackstone jointly announced the establishment of the "AI XPV Platform." The goal is ambitious: to support the deployment of over 20GW of global AI computing power by 2028.
But behind the grand vision lies a core mechanism that is a very specific financing transaction.
The first deal was worth $35 billion, with Anthropic as the client, aiming to build over 1GW of AI computing power for it.
The operation works like this: Atlas SP Partners, a structured finance division under Apollo, establishes a bankruptcy-remote Special Purpose Vehicle (SPV). This SPV raises funds by issuing debt, uses this money to purchase AI server racks equipped with Broadcom's custom XPU chips, and then delivers these racks to Anthropic on a 5-year lease. Anthropic pays rent each period, and the rent flows back to service the SPV's debt.
Simply put: Apollo puts up money to create a shell; the shell buys Broadcom's chips and leases them to Anthropic.
This structure benefits all three parties. Anthropic doesn't need to come up with $35 billion upfront to buy hardware; the equipment stays off its balance sheet, which is beneficial for its upcoming IPO—in fact, Anthropic concurrently completed a $65 billion Series H equity financing round, valuing it at $965 billion, with the two capital structures independent.
Apollo and Blackstone gain a long-duration asset with contractual cash flows, perfectly suited for the investment needs of their insurance funds (like Apollo's insurance subsidiary, Athene).
Broadcom secures a massive chip order.
But there's a key element: Who guarantees the safety of this debt?
Broadcom's Guarantee
The debt issued by the SPV is divided into three tranches: a $6 billion A1 tranche (sold to banks at a rate of U.S. Treasuries plus 100 basis points), a $24 billion A2 tranche (5.75% coupon, sold to institutional investors like Apollo's Athene), and a $4.5 billion B tranche (8.5% coupon, issued at a discount).
The combined ~$30 billion of senior debt in the A1 and A2 tranches is backed by a Residual Value Guarantee from Broadcom. The B tranche has no Broadcom guarantee.
The mechanism of the "Residual Value Guarantee" is this: If Anthropic stops paying rent (defaults), the SPV will first try to sell those AI server racks to repay the debt.
But chips are not airplanes—a Boeing 737 still has a mature secondary market for sale or re-lease even if an airline goes bankrupt, with relatively stable residual value.
AI chips are different. Technology iterations are extremely fast; today's top-of-the-line configuration may be obsolete tomorrow, and there is currently no mature secondary market for AI chips to validate residual value assumptions.
Therefore, if the proceeds from selling the chips are insufficient to repay all principal and interest of the senior debt, Broadcom covers the shortfall.
This is the essence of an RVG—a shortfall guarantee.
It has two triggering conditions: Anthropic defaults, AND the chip residual value falls below the guarantee threshold. Both conditions must be met for Broadcom to pay.
Under normal circumstances, as Anthropic pays rent on schedule and the debt amortizes, Broadcom's exposure decreases over time, reaching zero at the end of the 5-year term.
Broadcom disclosed this arrangement in its 10-Q filing for the period ended May 2026. The wording was quite restrained, not even naming Anthropic or Apollo, only mentioning "an arrangement with an investor partner," "providing backstop support for a customer's lease obligations," with a "maximum exposure of $29 billion."
This is an off-balance-sheet contingent liability, not recognized on the balance sheet.
If the story ended here, a $29 billion contingent exposure, while not small, wouldn't be fatal for a company generating over $20 billion in annual free cash flow and holding an A- investment-grade rating.
The problem is, this is just the first deal.
From $35B to $370B
The goal of the XPV Platform is not to do one $35 billion deal and stop, but to support 20GW of computing power by 2028—the first deal only covers a little over 1GW.
Bank of America's model assumes the platform expands at a pace of 2GW per quarter. Each new deal follows the same SPV structure, with Broadcom providing the same RVG for the senior debt. As new deals keep launching while old debt hasn't fully amortized, the cumulative RVG exposure Broadcom bears at any given point continues to climb.
Curcuruto's calculation result: By mid-2029, when the platform reaches 20GW scale, the cumulative principal amount of senior debt Broadcom faces could reach $370 billion—with about $150 billion in new issuance in 2027 alone. The corresponding maximum theoretical loss: $42 billion at a 100% default rate, $10.5 billion at a 25% default rate.
This number is, of course, the product of an extreme stress test, not debt Broadcom has already signed off on, nor is it a "$370 billion loss."
But it reveals a structural problem: If the platform expands as planned, Broadcom's off-balance-sheet contingent liabilities will balloon at a rate far exceeding its revenue and free cash flow growth.
And things are already accelerating. In early August, according to Bloomberg, Blackstone was already in talks with investors for a second financing package of at least $36 billion, also for Anthropic to lease Google's Ironwood TPUs. If finalized, the two deals would total ~$71 billion, and Broadcom's corresponding RVG exposure would double again.
S&P Global Ratings issued a warning as early as June 11th. It rated the $4.5 billion B tranche of the first XPV deal, which lacks a Broadcom guarantee, as "credit negative," stating the overall arrangement has a "moderate negative impact" on Broadcom's credit profile—although it maintained the A- rating, this characterization itself is a signal.
According to Bank of America, S&P currently tends to treat RVG directly as debt; if the platform continues to expand, whether S&P will adjust its framework, or even impact Broadcom's investment-grade rating, is one of the market's most sensitive uncertainties.
Why is a Chip Company Providing Financial Guarantees?
XPV causes unease not just because of the large numbers, but because it changes how people understand Broadcom as a company.
Before XPV, Broadcom was a semiconductor company designing chips and collecting design fees and royalties. Its business model was known for being asset-light and high-margin.
But in the XPV structure, Broadcom plays two roles simultaneously: it is both the equipment supplier (selling chips to the SPV) and the guarantor of the financing (backstopping the SPV's debt).
This dual role creates a subtle conflict of interest. The more chips Broadcom sells, the larger the platform grows, the better its revenue growth looks—but at the same time, its off-balance-sheet guarantee exposure also expands in sync.
In a sense, Broadcom is using its own investment-grade credit to swap for the credit risk of startups like Anthropic, in order to sell chips.
This is precisely the core of the "circular financing" criticism. The common feature of these arrangements: using private credit and SPVs to turn computing power into an asset class that can be underwritten, much like Wall Street packaged residential mortgages into securities two decades ago.
The real underlying risk is: Are the assumptions about AI chip residual value valid?
Chips Are Not Airplanes
Aircraft financing has operated for decades because a Boeing or Airbus aircraft has a useful life of 20-30 years, a mature secondary market, a large pool of potential buyers, and residual value curves validated by decades of data.
Financial institutions dare to provide billions in aircraft leasing finance to airlines because they know that even if the airline fails, the plane itself can be resold or re-leased.
AI chips are not that kind of asset. Nvidia GPU architecture cycles are about 1-3 years, and Broadcom XPU iteration rhythms are similar.
How much will chips bought for $35 billion today be worth in 5 years? No one can give a data-backed answer because a large-scale secondary market for AI chips simply does not exist yet.
Bank of America's model assumes chip prices decline 20% annually, with an additional 25% price shock upon default. Whether this assumption is conservative or optimistic is currently unverifiable.
In a typical real estate lease, the building preserves value, but the semiconductors at the heart of AI data centers rapidly depreciate due to fast-paced technological iterations.
The Bond Market is Already Voting with Its Feet
The credit market's reaction to XPV came earlier and was clearer than the stock market's.
Since the XPV announcement in June, Broadcom's bond spreads have widened by 30-45 basis points relative to similarly-rated semiconductor peers like Texas Instruments and Qualcomm.
In other words, bond investors are already demanding extra compensation for the risk associated with Broadcom's XPV guarantees.
The spreads on Broadcom's benchmark bonds maturing in 2036 (4.95%) and 2056 (5.7%) reached 105 and 118 basis points, respectively, significantly wider than non-AI semiconductor peers.
What is the Market Afraid Of?
Not that AI is failing.
Broadcom's AI chip revenue last quarter soared 143% year-over-year, with orders backlogged over two years; few question the demand side.
The market is afraid of something else: When the funding scale needed for AI construction reaches the trillion-dollar level, when chip companies start using their own credit to backstop their customers' leases, when off-balance-sheet contingent liabilities balloon much faster than revenue—how much risk in this chain is priced in, and how much is ignored?
With XPV, Broadcom found a shortcut to rapidly deploy AI computing power. But the price of the shortcut is that it is transforming from a light-asset chip design company into a "quasi-guarantor" for the semiconductor industry—using its A- rated credit to provide the final safety net for the entire financing chain of AI infrastructure construction.
This article is from the WeChat public account "Wall Street Insights", author: Yue Ming
