From CoreWeave to Anthropic, chip-collateralized financing is reshaping the AI debt market

The prototype of this market began in 2023 with a $2.3 billion private credit loan that Blackstone provided to CoreWeave, becoming the first landmark transaction of chip-collateralized debt.

Today, Broadcom is in negotiations with Wall Street institutions to arrange $60 billion in chip financing for Anthropic, SpaceX is reportedly seeking $40 billion in financing for Nvidia chips, and Nvidia itself has announced a $500 billion cooperation framework with six major lenders.

JPMorgan analysts estimate that total chip spending before 2030 may exceed $3 trillion, most of which will rely on debt financing.

However, this debt expansion also brings risks that cannot be ignored. Because most transactions are private in nature, investors lack convenient exit channels; cross-exposure among chipmakers, cloud computing giants, and emerging cloud service providers is highly concentrated.

CoreWeave's corporate-level junk bonds have recently fallen sharply, with some dropping to 80 cents on the dollar and yields climbing to about 15%, reflecting deep market doubts about the commercial returns of AI.

A single financing deal creates a new paradigm

The starting point of this revolution was the Blackstone Tactical Opportunities team's cautious assessment of an unfamiliar concept. According to CoreWeave co-founder Brannin McBee, who participated in the due diligence at the time, Blackstone executives directly asked back then: "What is a GPU?" "Is AI real?"

After receiving satisfactory answers, Blackstone led the completion of a $2.3 billion private credit loan to CoreWeave in 2023, setting a precedent for chip-collateralized debt financing.

The core logic of this structure is: establish a special purpose vehicle (SPV) to purchase and hold chips, repay debt with cash flows generated from target customer contracts, and match the debt term with the chips' five- to six-year depreciation cycle, thereby locking hardware depreciation risk within a predictable range.

CoreWeave's model was born from commodity market experience—similar to the development model in the liquefied natural gas sector of "sign contracts first, then finance and build plants." The first transaction was collateralized by a Microsoft contract, and at that time the financing cost was as high as a mid-teens interest rate level. McBee admitted that this was "the most difficult due diligence of my career."

As CoreWeave's customer lineup continued to expand, its financing costs declined accordingly. In March this year, the company completed $8.5 billion in debt financing supported by a Meta contract, with the borrowing rate dropping to about 6%—Wall Street priced the risk according to Meta's high rating rather than CoreWeave's own junk rating.

Chip costs far exceed data centers

To understand the scale of this financing revolution, one must first understand the magnitude of chip spending. The cost of chips is two to four times that of the data centers that house them, yet they depreciate fully within five to six years, creating a massive financing gap that traditional capital markets find difficult to absorb independently.

Take AI cloud service provider Nscale's flagship project in West Virginia as an example: the supporting power plant is expected to cost $6.5 billion, the data center and equipment $18 billion, and the GPU procurement cost as high as $47 billion.

Investment-grade cloud computing giants such as Alphabet's Google, Microsoft, Amazon, and Meta can cover these hardware costs with ample cash or cheap bond financing, but OpenAI, Anthropic, and a host of emerging cloud service providers do not have such balance sheet strength.

OpenAI Chief Business Officer Mike Liberatore admitted: "The scale of infrastructure construction is unprecedented. No playbook tells you how to go from zero to artificial general intelligence in the traditional way under extremely capital-intensive conditions."

To this end, the industry has transplanted mechanisms from project finance and structured finance: isolating the risk of each transaction through SPVs, borrowing based on the quality of the transaction itself and the customer's credit rather than the issuer's own rating, often without being recorded on the corporate balance sheet.

Chipmakers step in directly to endorse

The deep involvement of Broadcom and Nvidia has pushed this market to a new magnitude. The two companies provide "backstop" commitments for related debt transactions:

Once the borrower defaults, the chips will be re-leased or sold to repay the remaining debt, and the chipmakers are responsible for filling the gap. This mechanism enables senior bonds issued by SPVs to be linked to the chipmakers' investment-grade ratings, significantly lowering financing costs.

Apollo and Blackstone recently jointly completed a $35 billion "Big Sky" transaction to provide support for Anthropic. This is also the first transaction to land on the "AI XPV" platform jointly built by Broadcom, Apollo, and Blackstone—Bank of America strategists estimate that if the platform achieves 20 gigawatts of financing capacity, it may require $235 billion in backstopped debt by mid-2029.

In this Anthropic transaction, the two senior bonds backstopped by Broadcom received investment-grade ratings, with the largest tranche at a 5.75% interest rate; the riskier $4.5 billion subordinated tranche was priced at 8.5%—still far below Anthropic's standalone financing cost as an unrated private company.

In August this year, Nvidia further assembled Apollo, Blackstone, BlackRock, Brookfield, KKR, and Goldman Sachs to jointly provide financing sources for AI infrastructure, with Nvidia itself committing to backstop up to 25% of the share.

From private placement to syndicated market

To achieve true scale, chip financing must break through the boundaries of the private market and enter the syndicated loan and even investment-grade public bond markets—the latter being one of the largest and lowest-cost capital pools in the world.

CoreWeave took the lead this year in completing chip-backed syndicated leveraged loans, with both transactions receiving solid subscriptions from institutional investors and trading in the market near par, in sharp contrast to the same company's corporate bonds.

Rich Myers, global head of structured products and private capital markets at Morgan Stanley, said: "These financings are directly linked to the most valuable asset in the market, namely the computing power generated by chips. A large body of evidence shows that demand for computing power will continue to grow and exceed supply."

Chris Edson, head of global asset allocation at Apollo, said the traditional public debt market cannot single-handedly meet the financing scale of digital infrastructure, and requires participation from multiple types of capital such as project finance, private placements, insurance funds and commercial real estate. Obtaining an investment-grade rating is the key to unlocking the largest capital pools such as insurance funds.

Risk accumulation cannot be ignored

However, as capital accelerates its inflow, risks are accumulating in tandem.

Jason Pompeii, senior director at Fitch Ratings, pointed out that the current scale of vendor financing is unprecedented and risks are highly concentrated. "This only increases the possibility that once AI ultimately disappoints, the entire supply chain will be hit."

The private nature of the transaction structure raises doubts about liquidity. Because most chip financing transactions do not circulate in the public market, investors find it difficult to exit conveniently, while the transactions deeply bind a small number of chip manufacturers, cloud computing giants and emerging cloud service providers, with highly concentrated cross-exposure. Dave Snyderman, managing director of alternative credit and fixed income at Magnetar, warned: "The giants will only get bigger. Only they can obtain financing channels."

The legal validity of customer contracts is also in question. Akhil Bansal, head of asset-backed financing at Carlyle, said bluntly: "Everyone assumes these contracts are unbreakable, but what I focus on is the practical level—if only seven institutions provide all the credit support for GPUs, whose hands is the leverage really in?"

Laura Parrott, head of private fixed income at Nuveen, said she is currently holding off on participating in large investment-grade chip transactions and will reassess after the market structure becomes more mature. "This is a brand-new structure, and once some kind of shock occurs, we do not yet know how the outcome will unfold."

In addition, there are macro-level variables: local opposition in Texas has led to a suspension order for new projects, Oracle's project in New Mexico has been forced to delay payments due to regulatory obstacles, and OpenAI and Anthropic are also discussing slowing the pace of deployment over safety issues.

Moving toward a "computing power asset class"

Despite the challenges, market participants remain optimistic about the long-term potential of this paradigm.

John Watson, senior managing director at Blackstone's tactical opportunities team, compared the current moment to the early years of the automotive industry: those who financed the Ford Model T back then also faced the question "will people really drive?" Today, auto finance has developed into a large-scale, mature market.

At the level of chip market evolution, early research and actual usage data show that the actual service life of chips is exceeding the years reflected in current accounting treatment, which will improve their financeability. Karen Fang, global head of infrastructure and sustainable financing at Bank of America, expects that semiconductor valuation tools similar to used-car price guides will gradually emerge, thereby spawning a more active chip leasing and hedging market.

Watson's ultimate vision is to transform a unit of computing power into an asset class that can be traded and financed globally—"we understand the economic value of computing power, and therefore we finance it."

Markets carry risk, and investment requires caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation or needs of individual users. Users should consider whether any opinions, views or conclusions in this article are suitable for their particular circumstances. Investment based on this is at the user's own risk.