Broadcom Drops 6% Overnight: The Next Hurdle for AI Is Financing Costs

On August 14, Broadcom’s stock closed at $393.08, down 5.92% for the day, wiping out approximately $118 billion in market capitalization. During intraday trading, it fell as low as $388.50.

The trigger for the decline was a credit report by Bank of America analyst Tom Cucuruto. The report estimated that if the financing platform Broadcom built for its AI chip customers expands to a scale of 20 gigawatts as planned, its guaranteed exposure could reach $370 billion by mid-2029.

To investors, the $370 billion figure, linked to debt, constituted a signal requiring immediate repricing. However, this number does not represent debt that Broadcom needs to repay.

The real risk in the market is not a subprime-style collapse—AI chip residual values are still rising, but climbing financing costs are quietly eroding profit margins across the entire supply chain. This is a compression of valuation, not the end of the cycle.

One Number, Two Calculations

Where did this $370 billion come from? We must look back to XPV, the financing platform jointly established by Broadcom, Blackstone, and Apollo. The platform operates as follows: institutional investors purchase custom AI chips from Broadcom and then lease them to customers. The first transaction, valued at $35 billion and corresponding to approximately 1 gigawatt of compute capacity, had artificial intelligence company Anthropic as the lessee. Through this arrangement, the debt generated from purchasing the chips remains on the platform’s balance sheet and is not recorded as Anthropic’s liability.

Broadcom’s role in this structure is to guarantee the leases. If a customer stops paying rent, Broadcom will take over the equipment or resell it. In its latest 10-Q filing, Broadcom disclosed that for the first transaction, even in an extreme scenario where all customers default and the equipment has zero residual value, its maximum exposure would be $29 billion.

The $370 billion calculated by Bank of America represents the cumulative nominal cap of Broadcom’s guarantees after the platform expands to 20 gigawatts. This is a guarantee cap, which is distinct from debt awaiting repayment.

The same report also provided loss estimates: even with a 100% customer default rate, Broadcom’s actual losses would be approximately $42 billion; at a more realistic 25% default rate, losses would be around $10.5 billion. In comparison, Broadcom’s free cash flow after dividends in 2027 is projected to reach $85 billion.

A guarantee with a loss cap capped at $42 billion was priced by the market as $370 billion in debt. This is the logic behind Friday’s 6% drop.

This is not the first time this year that Broadcom has seen a divergence between its performance and its stock price. In July, the company raised its AI revenue guidance by over 200%, yet its stock still closed lower on the day.

The Fundamental Difference in Collateral

The market’s initial reaction was to analogize this structure to the 2008 subprime mortgage crisis. The transmission chain of 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 placed into special purpose vehicles (SPVs) as collateral to secure financing based on their expected cash flows—a tool Wall Street calls "Compute Collateralized Obligations" (CCOs). Compute cloud service provider CoreWeave issued the first such instrument in May this year, with a size of $3.1 billion, with underlying borrowers including OpenAI and Cohere.

Structural similarities led some market participants to directly characterize it as a replica of Collateralized Debt Obligations (CDOs).

Michael Burry, the prototype for "The Big Short," publicly increased his AI-related short positions this week and warned of a market bubble. On August 10, NVIDIA announced the establishment of a financing platform exceeding $500 billion with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR. The following day, its stock fell 2.8%, wiping out approximately $70 billion in market capitalization in a single day, while its five-year Credit Default Swaps (CDS) rose nearly 90% year-to-date.

The bears’ logical chain is complete. However, this analogy ignores a fundamental difference: what exactly constitutes the underlying asset of subprime loans.

The foundation of subprime loans was individuals lacking repayment ability. Upon default, the collateral was real estate that was depreciating and lacked liquidity.

The foundation of compute power is chips. NVIDIA CEO Jensen Huang provided a set of data in response to questions about "circular financing": the one-year lease price for H100 chips rose from $1.70 per GPU hour in October 2025 to $2.35 in March 2026; A100 chips that have been in service for six years are still running in production environments.

Collateral that is increasing in price, having its lifespan continuously extended by the CUDA software ecosystem, and can be transferred across customers and workloads, is fundamentally different from the underlying assets of the subprime crisis. This is also why NVIDIA dares to promise to cover up to 25% of each loan based on residual value.

The Risk Lies Not in Leverage, But in Profit Margins

However, this does not mean that this financing architecture poses no risk. The true area of concern lies not in "whether it will explode," but in "rising cost of capital."

An overlooked fact is that AI-related bond issuance reached $344 billion by early August this year, more than $200 billion higher than the full year of 2025. However, the vast majority of these funds were acquired by giants with investment-grade ratings such as Meta, Alphabet, and Amazon. Small and medium-sized AI companies lacking investment-grade ratings have instead been marginalized in the bond market. The cost of CoreWeave’s most recent loan was 5.5 percentage points above the benchmark rate, with a yield exceeding 9%.

The significance of financing platforms lies in providing another financing channel for customers who cannot access low-cost funding. The demand is real, and the cost is equally real—debt costs are rising.

A deeper issue concerns the sustainability of capital expenditures. The ratio of capital expenditures to operating cash flow for hyperscale cloud providers has risen from about 30% in 2022 to about 60% in 2025. According to market consensus, it will touch 100% in 2026. In other words, every subsequent unit of marginal compute investment will need to rely on debt or equity financing.

At the same time, signs that the growth rate of capital expenditures has peaked are emerging. While the total volume continues to hit new highs, 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, without waiting for growth to reach zero.

The credit market has reacted ahead of the stock market. NVIDIA’s five-year CDS has doubled since late May, jumping nearly 6 basis points in a single day on August 10 when the financing platform was officially announced. Bond traders have begun to price risk for the new asset class of "compute collateral," while the stock market only responded on Friday. This time lag indicates that the market is facing a new variable being seriously priced in, not just a simple fluctuation in sentiment.

Therefore, judgments on this event should be layered. The probability of systemic default is low because the underlying assets possess cash flow, residual value, and price appreciation potential; however, "rising cost of capital" is a happening reality, which will gradually erode valuations through profit margins. This is a compression of valuation, not the end of the cycle.

The dividing line lies in the default rate. If the default rate is 25%, a $10.5 billion loss is manageable for Broadcom; if the default rate reaches 100% and the equipment cannot be disposed of, a $42 billion loss would consume nearly half of Broadcom’s free cash flow—that would be a problem on another level.

Sliding from "valuation compression" to "breakage" requires two conditions to occur simultaneously: an actual default by a large borrower, and the mortgaged chips depreciating to the point where no one buys them at the same time. Regarding the former, Anthropic and OpenAI are far from investment-grade ratings and have a high dependence on refinancing; regarding the latter, the secondary market for chips, still-rising lease prices, and the life-extending capability of the CUDA ecosystem are all supporting residual values. The probability of both conditions triggering simultaneously is far lower than the panic implied by Friday’s 6% drop.

Who Is Being Repriced, Who Is Being Wrongly Punished

Following this logic, Friday’s market performance already provided its own answer.

The entity itself was the first to be repriced. Broadcom fell nearly 6% in a single day, while the iShares Semiconductor ETF fell only 0.7%, and Intel fell 2%. This indicates that the market was separately repricing Broadcom’s balance sheet, unrelated to the trend of the entire semiconductor sector.

Those wrongly punished were the cloud providers. On August 11, the day NVIDIA officially announced the financing platform, Google fell 3.84%, marking its largest single-day drop in six months, Amazon fell 2%, and Apple, Microsoft, and Broadcom all fell more than 1%. These companies have robust balance sheets, and the financing platform brings them an additional source of funding, which hardly constitutes a risk. Their stock price declines were more likely due to emotional selling after the market conflated "AI capital" with "AI debt."

The true beneficiaries are asset management firms. Apollo rose 6.26% on August 11, and KKR rose 6.88%. Financing platforms are their business; the larger the scale, the higher the management fee income. The market clearly distinguished "who is bearing the risk" from "who is collecting fees" through capital flows.

Another clue from the same day provided corroboration: Applied Materials’ third-quarter revenue and profit both hit record highs, yet its stock still fell 5% after hours. Strong performance coupled with a stock price decline reflects that the market is recalculating the risk premium for every company in the AI supply chain, even if their financial reports are impeccable.

If we return this logic to the market dynamics, the observation sequence should be: on the night the news was released, first focus on the CDS and stock prices of Broadcom and NVIDIA themselves, as they are the direct carriers of guarantee exposure; the next day, watch whether asset management stocks like Apollo, KKR, and Blackstone can continue their rise; in the following days, observe the performance of cloud providers and second-tier AI chip companies to determine whether their decline was a wrongful punishment or the beginning of sentiment transmission.

The deeper implication is that the 6% drop currently priced in by the market represents a one-time downward revision equating "guarantees" with "debt." But what truly needs to be gradually digested is a more persistent variable—when financing becomes the primary source of marginal compute capacity, the profit margins of the entire AI supply chain will have a portion of interest costs accounted for. This erosion will not be completed in a day, nor will it end in a day; it will constitute a more lasting pricing force than any single news item in the coming quarters.

Clarity in the Next Earnings Report

Subsequent observations should 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 stated it tends to treat such residual value guarantees as debt. If Broadcom proactively increases disclosure, it indicates it is treating this credit issue prudently; if it remains vague, it constitutes a true negative factor.

Second, the first actual fundraising of NVIDIA’s $500 billion financing platform. A memorandum of understanding is not capital; only actually implemented financing can prove the viability of this model. The smoothness of fundraising and the interest rate levels will directly determine whether the narrative of "compute as collateral" can continue.

Third, the capital expenditure guidance of the four major cloud providers for the next quarter. If the growth inflection point evolves from "peaking" to "declining," the judgment in this article will need to be revised—