AI infrastructure financing is undergoing a structural shift, and the market is beginning to price in this change. On August 14, shares of Broadcom closed over 5.9% lower, falling as much as 7% during the session. This sell-off was not driven by a sudden deterioration in earnings, but by investors reassessing the credit risks embedded in Broadcom's "seller guarantee" model, facilitated through its AI XPV financing platform.
At the same time, NVIDIA is pushing forward with a significantly larger, similar initiative. It is reportedly collaborating with Wall Street institutions including Apollo, Blackstone, BlackRock, and Goldman Sachs, aiming to mobilize over $500 billion in third-party capital through a computing financing platform. The moves by these two AI chip giants signal a new phase in AI infrastructure buildout: as hyperscale cloud providers' capital expenditures approach their financial limits, chip sellers are stepping in. By providing guarantees for Special Purpose Vehicles (SPVs), they are securitizing expensive computing assets and leveraging private credit market funds. This has raised a critical question for the market: is this model an orderly extension of AI capital spending, or a new form of financial leverage risk?
Why Are Hyperscalers Pushing Financing Innovation?
According to a Barclays research report, the direct catalyst for this new wave of AI financing demand is that hyperscale cloud providers are nearing their natural capital expenditure limits. Barclays estimates that the combined capital spending of the five major hyperscalers—Amazon, Google, Meta, Microsoft, and Oracle—will exceed their operating cash flow by 2026. The "funding gap" is projected to widen to approximately $210 billion by 2027 and grow further in 2028. The report notes that these hyperscalers face real constraints on debt levels and power agreements, with bond issuance already having expanded significantly. For some companies, the ratio of capital spending to operating cash flow has already surpassed 100%.
Concurrently, AI labs are experiencing extremely strong revenue growth. Barclays estimates that the annualized recurring revenue (ARR) of AI labs could exceed $200 billion by the end of 2026, roughly 2 times and 5 times higher than the previous forecasts for OpenAI and Anthropic, respectively. The tension between surging demand for computing power and the tightening capital budgets of cloud providers has created the conditions for new financing structures. In this context, an SPV structure that separates the data center "shell" from computing assets has begun to emerge. In 2026, the cost to build a 1GW data center is estimated at about $15 billion for non-computing assets and $35 billion for computing assets. The latter is larger, depreciates faster, and carries a higher risk of obsolescence, making it the core issue this new securitization structure aims to solve.
Broadcom's XPV Platform: The Chip Seller Underwriting Private Credit
Broadcom's AI XPV Platform, formally launched in June, was established in partnership with Apollo and Blackstone. It has an initial capital plan of $35 billion, with a target of supporting over 20GW of AI computing capacity by 2028. As described in the Barclays research, the basic logic of the financing structure is as follows: AI labs like Anthropic, or emerging cloud providers (neoclouds), purchase TPU computing units designed by Broadcom for Google and manufactured by Taiwan Semiconductor Manufacturing Company. These computing assets are then placed into an SPV. Broadcom provides an approximately 85% guarantee on the senior notes issued by the SPV, enabling Blackstone or other financial institutions to provide investment-grade financing. Anthropic then pays the SPV for computing power, with neoclouds like Fluidstack handling cluster operations and management.
This structure offers clear strategic value for Broadcom: customers can deploy computing power quickly without large upfront capital expenditures, while Broadcom can leverage financial institutions' capital to scale XPU market penetration and grow its AI chip revenue. However, by providing guarantees as a seller, Broadcom's balance sheet will also bear potential contingent liabilities as the platform expands. Barclays estimates that Broadcom's cumulative guarantee exposure could approach $739 billion by 2028.
NVIDIA's $500 Billion Plan: Similar Logic, Larger Scale
As Broadcom's XPV platform drew market attention, NVIDIA announced its own, much larger, similar plan this week. NVIDIA stated it has signed a memorandum of understanding with Apollo, Blackstone, BlackRock, Goldman Sachs, Brookfield, and KKR to jointly establish an AI computing infrastructure financing platform. The goal is to mobilize over $500 billion in third-party capital, corresponding to about 8 to 10 GW of computing power deployment. According to the Barclays research, NVIDIA's platform structure is similar to Broadcom's: neoclouds purchase GPUs and inject them into an SPV, with NVIDIA providing a guarantee of no more than 25% on the SPV's senior debt to lower financing costs. End users like OpenAI then lease GPU computing power from the neoclouds. A key difference from Broadcom is that while providing the guarantee, NVIDIA retains a share of the GPU rental revenue above a preset hourly rate.
Barclays expects the new AI securitization market to grow rapidly, potentially accounting for about 20% of total industry capital spending in 2027 and, if successful, nearly 50% by 2028. Meanwhile, related guarantee exposures are likely to gradually appear in the financial filings of companies like Google, NVIDIA, and Broadcom. Google's Q2 2026 10-Q filing disclosed purchase obligations of $811 billion, of which Barclays believes roughly half could be attributable to data center guarantee exposure.
Market Repricing: Guarantee Size, Chip Residual Value, and Tenant Concentration
The sharp decline in Broadcom's stock price reflects the market's reassessment of the credit risks in this model. Analyst Tom Curcuruto from Bank of America noted that the XPV platform currently has high tenant concentration, with the initial transaction primarily relying on Anthropic. Although OpenAI may become a future client, other tenants have not been clearly identified. If the platform's tenants are too concentrated, a debt repayment problem with a single core client could put significant pressure on the related SPV.
Additionally, the uncertainty around the residual value of custom AI chips poses a potential risk. Unlike assets like aircraft or servers that have mature secondary markets, Broadcom's custom XPUs lack sufficient liquidity. If a client defaults, the price and speed at which these chips can be disposed of are highly uncertain, directly impacting the financing safety margin of the SPV. The Barclays research team offered a more positive assessment of these concerns. The report argues that there is a high degree of information transparency and incentive alignment among AI chip sellers, manufacturers, cloud providers, and AI labs, making the probability of a large-scale default relatively limited. It also notes that the high versatility of NVIDIA's GPUs means that even if demand weakens in one area, computing resources can be reallocated to other strong-demand scenarios.
However, the deeper question weighing on the market is this: if sustaining the current AI capital spending boom requires increasing amounts of debt financing and seller guarantees, how solid is the real cash flow foundation underpinning this growth? The answer to this question will likely continue to influence the valuation of the AI chip sector.
Comments