A surge in financing for AI chips has given rise to a new type of off-balance-sheet guarantee structure, with a total value of $70 billion, leaving the bond market grappling with how to price this hidden risk. This unease emerged even before Nvidia's announcement of a $500 billion financing collaboration, as bond investors began to fret over approximately $70 billion in "ghost liabilities" sitting outside the balance sheets of major AI companies. These contingent liabilities, typically inconspicuous, could suddenly crystallize at the worst possible moment.

The vehicle for these liabilities is a structural arrangement known as a "Residual Value Guarantee" (RVG), which can amount to hundreds of billions of dollars. In essence, the mechanism sees Nvidia leverage its own strong credit rating to back customer financing, helping clients reduce their borrowing costs. The typical structure operates in three layers: a Special Purpose Vehicle (SPV) borrows money to purchase chips; the borrowing is supported by cash flows from contracts signed by the company using the technology; if that company stops paying, the assets are re-leased or sold to repay the remaining debt; if there is still a shortfall, the "guarantor" makes up the difference. The chip seller acts as the ultimate backstop in this "Residual Value Guarantee" framework.

For chip giants like Nvidia and Broadcom, this structure is a "good deal": it helps clients lower financing costs and expand sales, while the company itself records no debt on its books. Meta stated directly in its filings: "The probability of the RVG guarantor being required to make a payment is not considered high, and therefore no liability has been recorded to date." While "probability is not high," this assertion is increasingly difficult for the market to accept. Broadcom has extended this logic into chip financing. In a project codenamed "Big Sky," Broadcom provided a guarantee for a $35 billion debt transaction, where investors like Apollo Global Management and BlackRock funded the purchase of custom AI chips, which were then leased to Anthropic. This arrangement allowed the senior debt to achieve an investment-grade rating, thereby lowering financing costs.

Unlike data center deals spanning decades, chip financing cycles are shorter, typically amortizing over about five years to match the rapid depreciation of technology. This means the guarantee exposure narrows quickly over time, providing lenders with a relatively clear exit horizon. With Nvidia's entry, the scale could jump again. Nvidia CEO Jensen Huang posted on Platform X that the company might offer up to a 25% residual value support mechanism for relevant opportunities, assessed on a "case-by-case basis." "Our role is to help unlock a large pool of independent capital while maintaining disciplined risk exposure," he wrote. Nvidia stated that this collaboration aims to bring in external capital to alleviate the "circular financing" issue, where AI companies fund each other's product purchases. The six US investment institutions involved in the $500 billion financing include BlackRock and Goldman Sachs. Broadcom's AI XPV platform is an extension of the "Big Sky" deal, and according to estimates by Bank of America strategists, the platform could accumulate $370 billion in senior debt by mid-2029. This suggests the potential for off-balance-sheet guarantees to be far greater than currently seen.

Rating agencies have issued warnings. Moody's wrote in a report: "The main risk lies in the high density of such transactions occurring in a short period." Moody's further noted: "We believe that a significant increase in Broadcom's contingent obligations, even if its existing debt leverage remains low, would constrain Broadcom's financial flexibility and could put pressure on the company's credit profile." Moody's also pointed out that while Broadcom's guarantees on third-party leases "partially offset strong business advantages," more guarantees would have a negative impact. S&P Global Ratings characterized the residual value support provided by Broadcom as a "contingent debt-like obligation" and stated it would be included in adjusted debt calculations. Under US accounting standards, companies typically only record contingent liabilities on their balance sheets when a loss is "probable and reasonably estimable"; otherwise, disclosure in the footnotes to financial statements suffices. This is the institutional basis for these liabilities remaining off the books.

Bond investors' concerns center on one point: when will these off-balance-sheet contingent liabilities become real on-balance-sheet losses? DoubleLine portfolio manager Mariya Entina stated bluntly: "It feels like a way to game the system, trying to get favorable treatment from rating agencies to achieve the highest possible rating... We are entering an era of financial engineering. This is one of my concerns: when you engage in financial engineering, you are masking financial reality." CreditSights analysts, in a report, compared Nvidia's residual value support to "selling a put option." "It is pro-cyclical and amplifies the potential for boom and bust," they wrote. "During a boom phase, the guarantee costs almost nothing; but in a severe, sharp downturn, when customers default and hardware market values fall, it becomes most critical." TCW Global Co-Head of Credit Brian Gelfand remarked: "This is not ordinary investment-grade credit underwriting. It is far more complex. Given the off-balance-sheet nature, the tail risk is elevated."

Some believe concerns are overblown. Janus Henderson Investors Global Head of Multi-Sector and Corporate Credit John Lloyd argued that triggering the residual value support would require extreme conditions: "You would need to see a cliff-like drop in the growth rate of token usage, which is absolutely not what we are seeing." He also noted that these companies "are not trying to hide contingent liabilities, but rather to finance them." Supporters' logic is that chip demand will outstrip supply for years to come; the debt structure is designed to amortize fully over time, with the potential cost of residual value support diminishing accordingly; and the technology risk ultimately falls on large tech companies with sufficient cash to absorb losses. However, critics' counterargument is equally potent: these guarantees are precisely triggered during industry downturns, which is also when chip manufacturers themselves face earnings pressure. The synchronization of the guarantee and risk with the economic cycle is the core problem.