On August 14, Broadcom shares closed at $393.08, falling 5.92% in a single day and wiping out approximately $118 billion in market value, with the stock briefly touching $388.50 during the session.
The trigger was a credit report from Bank of America analyst Tom Kulkuruto. The report estimated that if Broadcom's financing platform for AI chip customers expands to 20 gigawatts as planned, the company's guaranteed exposure could reach $370 billion by mid-2029.
In the eyes of investors, the $370 billion figure, tied to debt, became a signal demanding immediate repricing. However, this number does not represent debt that Broadcom must repay.
The real risk to the market is not a subprime-style collapse—AI chip residual values are still rising—but the creeping impact of rising financing costs on the profit margins of the entire supply chain. This is a valuation compression, not a cycle ending.
One figure, two calculations
Where does the $370 billion come from? It traces back to the XPV financing platform co-established by Broadcom, Blackstone, and Apollo. The platform's operation is: institutions buy Broadcom's custom AI chips and then lease them to clients. The first transaction, worth $35 billion, corresponds to about 1 gigawatt of computing power, with the lessee being AI company Anthropic. Through this arrangement, the debt from purchasing chips remains on the platform's balance sheet, not on Anthropic's.
Broadcom's role is to guarantee the leases. If a client stops paying rent, Broadcom takes over the equipment or resells it. In its latest 10-Q filing, Broadcom disclosed that for the first transaction, even in the extreme scenario of a total client default and zero equipment residual value, its maximum exposure is $29 billion.
The $370 billion estimated by Bank of America is the cumulative nominal cap on Broadcom's guarantees if the platform expands to 20 gigawatts. This is the upper limit of guarantees, distinct from debt to be repaid.
The same report provides loss estimates: even with 100% client default, Broadcom's real loss would be about $42 billion; under a more realistic 25% default rate, the loss would be around $10.5 billion. In comparison, Broadcom's post-dividend free cash flow in 2027 is projected at $85 billion.
A guarantee with a loss cap of $42 billion is being priced by the market as $370 billion in debt. This is the logic behind Friday's 6% decline.
This is not the first time this year that Broadcom's performance has diverged from its stock price. In July, the company raised its AI revenue guidance by over 200%, yet the stock still closed lower that day.
Fundamental differences in collateral
The market's initial reaction was to compare this structure to the 2008 subprime crisis. The subprime crisis transmission chain was: people lacking repayment ability got loans to buy homes, and the risk was securitized and spread across the global financial system. Today, GPUs are placed into special purpose vehicles as collateral, with financing based on their expected cash flows—Wall Street calls this tool a "Compute Capacity Obligation" (CCO). Cloud computing service provider CoreWeave issued the first such instrument in May this year, worth $3.1 billion, with underlying borrowers including OpenAI and Cohere.
The structural similarity has led some market participants to directly label it a version of Collateralized Debt Obligations (CDOs).
Michael Burry, the protagonist of "The Big Short," publicly increased his AI-related short positions this week, warning of market bubbles. On August 10, NVIDIA announced a financing platform worth over $500 billion with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR. The next day, its stock fell 2.8%, losing about $70 billion in market value, and its five-year credit default swap (CDS) year-to-date increase approached 90%.
The bearish logic chain is complete. But this analogy ignores a fundamental difference: what is the underlying asset of subprime mortgages?
The underlying asset of subprime was people without repayment ability. Once defaulted, the collateral was depreciating, illiquid real estate.
The underlying asset of computing power is chips. NVIDIA CEO Jensen Huang, responding to "circular financing" questions, provided data: the one-year lease price for H100 chips rose from $1.70 per GPU hour in October 2025 to $2.35 in March 2026; the six-year-old A100 chip is still running in production environments.
A collateral asset whose price is rising, whose lifespan is continuously extended by the CUDA software ecosystem, and which can be transferred across clients and workloads, is fundamentally different from the underlying assets of the subprime crisis. This is why NVIDIA dares to promise a backstop of up to 25% on each loan, based on residual value.
Risk lies not in leverage, but in profit margins
This does not mean the financing structure is risk-free. The real concern is not "whether it will implode," but "the rising cost of capital."
An overlooked fact is that AI-related bond issuance has reached $344 billion as of early August this year, more than $200 billion above the full-year 2025 total. However, the vast majority of these funds are accessed by blue-chip companies like Meta, Alphabet, and Amazon with investment-grade ratings. Smaller AI companies lacking investment-grade ratings are being marginalized in the bond market. The cost of CoreWeave's latest loan was 5.5 percentage points above the benchmark rate, with a yield exceeding 9%.
The purpose of financing platforms like XPV is precisely to provide an alternative funding channel for these clients who cannot access low-cost capital. The demand is real, but so is the cost—debt costs are rising.
The deeper issue is the sustainability of capital expenditure. The ratio of capital expenditure to operating cash flow for hyperscale cloud providers has risen from about 30% in 2022 to around 60% in 2025, and consensus estimates it will hit 100% in 2026. In other words, every marginal compute investment thereafter will need to rely on debt or equity financing.
Meanwhile, the growth rate of capital expenditure is showing signs of peaking. Total volumes are still hitting new highs, but the "acceleration" of growth has turned negative. This is the most critical link in the 2008 logic: credit structures often break at the inflection point of growth, not waiting for growth to reach zero.
The credit market has already reacted ahead of the stock market. NVIDIA's five-year CDS has doubled since late May, and on the day of the August financing platform announcement, it jumped nearly 6 basis points in a single day. Bond traders are already pricing risk for this new asset class of "compute collateral," while the stock market only responded on Friday. This time lag indicates the market is facing a new variable being seriously priced, not just a simple emotional fluctuation.
Therefore, the judgment on this event should be layered. The probability of a systemic default is low, because the underlying assets have cash flow, residual value, and pricing power. However, "rising capital costs" is a happening fact that will gradually erode valuations through profit margins. This is a valuation compression, not a cycle ending.
The dividing line lies in the default rate. At a 25% default rate, the $10.5 billion loss is manageable for Broadcom; at a 100% default rate with unsellable equipment, the $42 billion loss would consume nearly half of Broadcom's free cash flow—a different order of magnitude entirely.
Sliding from "valuation compression" to "breakdown" requires two conditions to be met simultaneously: an actual default by a large borrower, and the pledged chips depreciating to the point of no buyer. For the former, Anthropic and OpenAI are still far from investment-grade ratings, with high refinancing dependency. For the latter, the secondary market for chips, still-rising lease prices, and the CUDA ecosystem's life-extending ability all support residual values. The probability of both conditions triggering simultaneously is far lower than the panic implied by Friday's 6% drop.
Who is being revalued, who is being unfairly sold
Following this logic, Friday's trading session has provided its own answer.
The first to be revalued was the entity itself. Broadcom fell nearly 6% on the day, while the iShares Semiconductor ETF fell only 0.7%, and Intel fell 2%. This indicates the market is independently repricing Broadcom's balance sheet, unrelated to the broader semiconductor sector.
Those unfairly sold were the cloud providers. On August 11, the day NVIDIA announced its financing platform, Google fell 3.84%, its biggest single-day drop in six months. Amazon fell 2%, while Apple, Microsoft, and Broadcom all fell over 1%. These companies have solid balance sheets, and the financing platform provides them with an additional source of capital, not a risk. Their stock declines were largely driven by emotional selling, as the market confused "AI capital" with "AI debt."
The true beneficiaries are asset managers. Apollo rose 6.26% on August 11, and KKR rose 6.88%. Financing platforms are their business; the larger the scale, the higher their management fees. The market's capital flows clearly distinguished between "who is bearing risk" and "who is collecting fees."
Another clue on the same day provided corroboration: Applied Materials reported record third-quarter revenue and profit, but its stock still fell 5% after hours. Strong earnings yet a declining stock price reflect the market recalculating risk premiums for every company in the AI supply chain, even if their financial reports are flawless.
If we place the logic back on the market, the order of observation should be: on the night of the announcement, first look at the CDS and stock prices of Broadcom and NVIDIA, as they are the direct carriers of guarantee exposure. The next day, watch whether asset manager stocks like Apollo, KKR, and Blackstone can continue their rally. Over the following days, observe the performance of cloud providers and second-tier AI chip companies to determine if their decline is an unfair sell-off or the beginning of sentiment contagion.
The deeper implication is that the market's 6% decline is pricing a one-time downward revision that equates "guarantees" with "debt." But what it truly needs to digest gradually is a more persistent variable: when financing becomes the primary source of marginal computing power, the profit margins of the entire AI supply chain will be burdened by an interest cost. This erosion will not be completed in one day, nor will it end in one day. It will constitute a more sustained pricing force over the next several quarters than any single piece of news.
Next earnings season will reveal the truth
For future observation, focus on three signals.
First, Broadcom's earnings report in early September. Will the company proactively clarify the guarantee scope of the XPV platform? Will rating agencies follow up with adjustments? S&P has already indicated a tendency to treat such residual value guarantees as debt. If Broadcom increases disclosure, it shows it is prudently addressing this credit issue; if it remains vague, it will be a genuine negative.
Second, the first actual capital raise for NVIDIA's $500 billion financing platform. A memorandum of understanding is not capital; only actual funding raised can prove the model is viable. The smoothness of the fundraising and the interest rate level will directly determine whether the "compute as collateral" narrative can continue.
Third, the capital expenditure guidance from the four major cloud providers in the next quarter. If the growth inflection point shifts from "peaking" to "declining," the judgment in this article will need revision.
Comments