Broadcom Inc. (AVGO.US) fell 6% overnight: The next hurdle for AI is the cost of financing.

On August 14, Broadcom Inc. (AVGO.US) closed at $393.08, down 5.92% for the day, with a market value evaporating by approximately $118 billion, dipping as low as $388.50 at one point during trading. The trigger for this drop was a credit report by Bank of America analyst Tom Kuruluto. The report estimated that if the financing platform set up by Broadcom Inc. for AI chip customers expands to a scale of 20 gigawatts as planned, its guaranteed exposure could reach $370 billion by mid-2029. For investors, this $370 billion is associated with debt, signaling a need for immediate repricing. However, this figure does not represent actual debt that Broadcom Inc. needs to repay. The real risk in the market is not akin to a subprime collapseAI chip residual values are still on the risebut the climbing financing costs are quietly eroding profit margins across the entire industry chain. This is a valuation compression, not the end of a cycle. One number, two algorithms To find the source of the $370 billion, we need to look back at the financing platform XPV jointly set up by Broadcom Inc., Blackstone, and Apollo. This platform operates by having institutions fund the purchase of custom AI chips from Broadcom Inc., which are then leased to customers. The first transaction amounted to $35 billion, corresponding to approximately 1 gigawatt of computing power, with the lessee being AI firm Anthropic. This arrangement allows the debt incurred from purchasing the chips to remain on the platform's balance sheet, not counted as liabilities for Anthropic. Broadcom Inc. serves as a guarantor for the leases. If a customer stops paying rent, Broadcom Inc. will take over the equipment or resell it. In its latest 10-Q filing, Broadcom Inc. disclosed that regarding the first transaction, even in an extreme scenario where clients default completely and the equipment's residual value is zero, its maximum exposure would be $29 billion. The $370 billion calculated by Bank of America is the nominal upper limit of Broadcom Inc.'s cumulative guarantees after expanding the platform to 20 gigawatts. This is a ceiling on guarantees, which is different from repayable debt. The same report also provided loss estimates: even with a 100% client default rate, Broadcom Inc.'s actual loss would be around $42 billion; with a more realistic 25% default rate, the loss would be about $10.5 billion. In contrast, Broadcom Inc.'s free cash flow after dividends in 2027 is expected to reach $85 billion. A loss limit capped at $42 billion in guarantees is being priced by the market as a $370 billion debt. This is the logic behind the 6% drop on Friday. This is not the first time Broadcom Inc. has diverged from its performance and stock price this year. In July, the company raised its AI revenue guidance by over 200%, yet the stock still closed down that day. The fundamental difference in collateral The market's initial reaction was to compare this structure to the subprime crisis of 2008. The chain of transmission in the subprime crisis was: individuals lacking repayment ability obtained loans to buy homes, and the risk was securitized and dispersed throughout the global financial system. Today, GPUs are packaged into special purpose vehicles as collateral, financing based on their expected cash flowsWall Street refers to this instrument as "computing power collateral certificates" (CCO). CoreWeave issued the first such instrument worth $3.1 billion in May this year, with underlying borrowers including OpenAI and Cohere. This structural similarity led some market participants to directly characterize it as a version of collateralized debt obligations (CDOs). Michael Burry, the protagonist of "The Big Short," publicly increased his short positions related to AI this week, warning of a market bubble. On August 10, NVIDIA Corporation announced the establishment of a financing platform worth over $500 billion in collaboration with Apollo, BlackRock, Blackstone, Bow Capital, Goldman Sachs Group, and KKR, causing a 2.8% drop in its stock the next day, evaporating about $70 billion in market value, with its five-year credit default swap (CDS) soaring by nearly 90% this year. Short sellers have a complete logical chain. However, this analogy overlooks a fundamental difference: what are the underlying assets of the subprime mortgages? The underlying asset of subprime is individuals lacking repayment ability. Upon default, the collateral is real estate that depreciates and lacks liquidity. The underlying asset for computing power is chips. NVIDIA Corporation's CEO Jensen Huang provided a set of data in response to questions about "circular financing": the annual leasing price of the H100 chip has risen from $1.70 per GPU hour in October 2025 to $2.35 in March 2026; the A100 chip, after six years, is still operational in production environments. A collateral that is appreciating, with a lifespan extended by the CUDA software ecosystem, and transferable across customers and workloads, is fundamentally different from the subprime crisis's underlying assets. This is why NVIDIA Corporation is able to boldly promise a maximum 25% backing for each loan based on residual value. The risk lies not in leverage, but in profit margins However, this does not mean that this financing structure does not pose risks. The truly concerning aspect is not "whether a crisis will erupt," but "the rising cost of funds." An overlooked fact is that, since the beginning of this year, the issuance of AI-related bonds has reached $344 billion by early August, over $200 billion higher than the total for 2025. However, the overwhelming majority of this funding has been secured by investment-grade rated giants such as Meta, Alphabet, and Amazon.com, Inc. Mid-sized AI companies lacking investment-grade ratings have been marginalized in the bond market. CoreWeave's most recent loan cost was 5.5 percentage points above the benchmark rate, with yields exceeding 9%. The significance of the financing platform lies in providing an alternative source of funding for these clients who cannot access low-cost financing. Demand is genuine, and so is the costthe cost of debt is on the rise. A deeper issue lies in the sustainability of capital expenditure. The capital spending of major cloud providers as a percentage of operating cash flow has risen from about 30% in 2022 to around 60% in 2025. According to market consensus, it is projected to reach 100% in 2026. In other words, every marginal input of computing power thereafter will need to rely on debt or equity financing. At the same time, the growth rate of capital expenditure is already showing signs of topping out. While total spending continues to hit new highs, the "acceleration" of its growth rate has turned negative. This is the most critical point in the 2008 logic: credit structures often break at inflection points of growth rates, without waiting for growth rates to hit zero. The credit market has already responded ahead of the stock market. NVIDIA Corporation's five-year CDS has doubled since late May; on the day of announcing the financing platform in August, it jumped nearly 6 basis points in one day. Bond traders have begun to price risks for the new asset class of "computing power collateral," while the stock market only responded on Friday. This time lag indicates that the market is facing a new variable that is being seriously priced, not just a simple emotional fluctuation. Therefore, the judgment of this event should be layered. The probability of systemic defaults is low because the underlying assets generate cash flows, have residual value, and are capable of appreciating; however, the "rising cost of funds" is a reality that is occurring, and it will gradually erode valuations through profit margins. This is a valuation compression, not the end of a cycle. The dividing line is the default rate. A 25% default rate resulting in a $10.5 billion loss would still be bearable for Broadcom Inc.; however, if the default rate reaches 100% and the equipment cannot be disposed of, a $42 billion loss would consume nearly half of Broadcom Inc.'s free cash flowthis would pose another level of problem. Transitioning from "valuation compression" to "breakdown" requires two conditions to be met simultaneously: a large borrower actually defaulting and the collateralized chips depreciating to the point where no one would buy them. On the former, Anthropic and OpenAI are still far from achieving investment-grade ratings and are highly reliant on refinancing; on the latter, the secondary market for chips, the still-rising leasing prices, and the longevity provided by the CUDA ecosystem all support their residual values. The probability of both conditions being triggered simultaneously is far lower than the panic implied by Friday's 6% drop. Who is repriced and who is wrongly killed According to this logic, Friday's market has already given its own answer. The first to be repriced was Broadcom Inc. It fell nearly 6% in a single day, while the iShares Semiconductor ETF fell only 0.7% during the same period, and Intel Corporation dropped by 2%. This indicates that the market is repricing Broadcom Inc.'s balance sheet independently, regardless of the performance of the entire semiconductor sector. The ones wrongly punished were the cloud providers. On the day NVIDIA Corporation announced the financing platform on August 11, Alphabet Inc. Class C fell 3.84%, marking its largest single-day drop in six months, while Amazon.com, Inc. declined by 2%, and Apple Inc., Microsoft Corporation, and Broadcom Inc. all dropped more than 1%. These companies have solid balance sheets, and the financing platform provides them with an additional source of funding, which cannot be classified as a risk. Their stock price declines stem more from market emotional sell-offs that conflate "AI funding" with "AI debt." The real beneficiaries are the asset management firms. Apollo rose 6.26% on August 11, and KKR climbed 6.88%. The financing platform is their business; the larger the scale, the higher the management fee income. The market clearly distinguished "who bears the risk" from "who collects the fees" through the flow of funds. Another clue from the same day supports this: Applied Materials reported record highs in both revenue and profit for its third fiscal quarter, yet its stock still fell by 5% after hours. Strong performance coupled with a stock price decline reflects that the market is recalculating the risk premiums for every company in the AI supply chain, even if their financial reports are impeccable. If we revert the logic back to the market context, the order of observation should be: on the night of the announcement, focus first on the CDS and stock prices of Broadcom Inc. and NVIDIA Corporation, which are the direct carriers of the guarantee exposure; the next day, monitor whether the asset management stocks such as Apollo, KKR, and Blackstone can continue to rise; in the following days, observe the performance of cloud providers and second-tier AI chip companies to determine whether their declines are unjustified or the start of emotional transmission. The deeper implication is that the market is currently pricing this 6% drop as a one-time correction equating "guarantee" to "debt." However, what it truly needs to gradually digest is a more enduring variablewhen financing becomes the primary source of marginal computing power, the profit margins across the entire AI supply chain will be treated as a portion of interest costs. This erosion will not be completed in one day, nor will it end in one day; it will constitute a more persistent pricing force over the coming quarters than any single piece of news. The next financial report will tell the tale Future observations should focus on three signals. First, Broadcom Inc.'s financial report in early September. Whether the company actively clarifies the guarantee terms of the XPV platform, and whether rating agencies will follow up with adjustmentsS&P has expressed a tendency to treat such residual value guarantees as debt. If Broadcom Inc. proactively increases its disclosures, it indicates a cautious approach to this credit issue; if it remains vague, it will constitute a real negative. Second, the first actual fundraising from NVIDIA Corporation's $500 billion financing platform. A memorandum of understanding is not funding; only actual confirmed financing can prove the feasibility of this model. The ease of fundraising and the level of interest rates will directly determine whether the narrative of "computing power as collateral" can continue. Third, the capital expenditure guidance from the four major cloud providers for the next quarter. If the growth rate inflection point evolves from "reaching a peak" to "declining," the judgment presented in this article will need to be revised This article is reproduced from "Wall Street Journal"; GMTEight editor: Yan Wencai.