Chipwrecked: How Nvidia’s Circular Deals Built an AI Debt Boom

The AI data center build-out rests on two things: Nvidia chips and borrowed money. That dependence has produced a financing machine in which Nvidia invests in cloud providers, those providers borrow against Nvidia chips, and the proceeds go back to Nvidia for more chips. The arrangement strengthens Nvidia’s grip on the AI compute market. It also creates risks for lenders, banks, and the broader economy. Nvidia has made more than 70 AI investments this year, and its backing has helped create a class of neoclouds, including CoreWeave, that are built on debt and rapid hardware refresh cycles.

Key facts

• Nvidia chips are the backbone of AI data centers and the collateral behind a fast-growing debt market.
• Neoclouds borrow against GPUs, often Nvidia hardware, then use the money to buy more chips and build capacity.
• CoreWeave pioneered GPU-backed loans, starting with a $2.3 billion facility and later adding $7.5 billion and $2.6 billion deals.
• Lenders accept loan-to-value ratios from 50 percent to 110 percent; one early CoreWeave loan carried a 14 percent floating interest rate.
• Private credit firms, which are less regulated than banks, are major lenders; banks also lend to private credit and directly to AI companies.
• Depreciation assumptions are disputed. Some investors argue that large cloud companies understate chip depreciation by $176 billion from 2026 to 2028.
• Nvidia’s cloud service agreements total $26 billion, with $22 billion due by 2031, and may act as a backstop for neocloud demand.
• Google’s TPUs and custom chips from Amazon, Microsoft, Meta, OpenAI, AMD, Broadcom, Huawei, ByteDance, and Alibaba are increasing competitive pressure.

The circular machine at the heart of the AI build-out

CoreWeave is the clearest example. The company started with a $2.3 billion GPU-backed loan in 2023, then added a $7.5 billion facility in 2024 and a $2.6 billion loan in July. Lenders have accepted loan-to-value ratios ranging from 50 percent to 110 percent, depending on the maturity and the quality of the contracts attached to the chips. One early CoreWeave loan carried a 14 percent floating interest rate in the third quarter of this year. The logic is simple: GPUs are treated like trucks or aircraft, assets that can be repossessed and resold if a borrower defaults. The problem is that no one agrees on how quickly a GPU loses value.

For Nvidia, the model has obvious advantages. A dollar of equity investment can support several dollars of chip purchases. The neocloud buys Nvidia hardware, puts it in a data center, and borrows against it. That borrowed money becomes revenue for Nvidia. The same dynamic gives Nvidia leverage over its largest customers. Its revenue is concentrated: in its most recent filings, two direct customers represented 21 percent and 13 percent of revenue in the first nine months of its fiscal 2026. Keeping a stable of neoclouds alive gives Nvidia alternative buyers and bargaining power.

Depreciation is the disputed variable

GPU-backed loans depend on assumptions about how long chips remain economically useful. Many companies depreciate GPUs over five or six years. Some researchers argue that is too long for Nvidia chips, which face rapid generational improvements. Google’s custom TPUs may last six years because they are designed for AI from the start. Nvidia’s Ampere chips, released in 2020, are still in use, and some operators say they are fully booked. Older chips can remain useful for inference, mining, or less demanding workloads. But newer chips are more efficient per operation, and power is a binding constraint. A company that can charge less per operation can win contracts, which forces neoclouds to keep upgrading.

The depreciation debate is not only an accounting issue. It affects the collateral behind billions of dollars of debt. If chips lose value faster than projected, borrowers may have to top off loans with more collateral or more cash. A prominent short-seller has argued that large cloud companies understate depreciation by $176 billion between 2026 and 2028. Whether or not that estimate is correct, the dispute shows how fragile the assumptions are. Lenders are making long-term loans against assets whose economic life is measured in a few years, not decades.

The economics can work in the best case. Industry estimates suggest that older V100 chips renting for $2 to $3 per hour can recoup their purchase price in four to seven months if fully utilized. Newer B200 nodes can cost more than $100 per hour and may pay back in about six months. But utilization is never guaranteed. Power, cooling, construction delays, and contract terms can all reduce returns. A data center that opens months late may lose part of the brief window when its chips earn maximum value.

Private credit and the banking link

Private credit firms are central to the GPU debt boom. They need to put investor capital to work, and AI offers high yields at a time when many traditional deals are scarce. Private credit has grown roughly tenfold since 2009, and it is less regulated than bank lending. The loans are often larger, later in line for repayment, and more expensive than bank debt. Banks are also connected to the system. They lend directly to AI companies and indirectly to private credit providers. As of June, banks had lent about $300 billion to private credit firms. That creates a channel through which stress in AI lending can reach regulated institutions.

The structure of some deals adds more opacity. Private credit firms can create special-purpose vehicles that allow companies to borrow without putting debt on their balance sheets or affecting their credit ratings. Data centers are also issuing asset-backed securities, and data center debt is being packaged into derivative products. A former Treasury official has compared these structures to the mortgage derivatives that amplified the 2008 financial crisis. Special-purpose vehicles also proliferated before that crisis because they hid how leveraged companies really were.

The scale of AI spending makes the financing more consequential. Analysts at Morgan Stanley expect $3 trillion in data center spending by the end of 2028. The tech sector has taken on more debt than it did during the dot-com bubble, according to Moody’s Analytics. In the 1990s, the exuberance was mostly in equity, so losses fell mainly on investors in failed companies. Debt changes the distribution of risk. If AI spending falters, the damage can spread to lenders, banks, pension funds, universities, family offices, and endowments.

Nvidia’s backstop problem

Nvidia has a strong incentive to keep neoclouds afloat. It is an investor in several of them, and their purchases support its revenue. It also has a history of stepping in when the market turns. During the crypto bust, Nvidia was left with more than $1 billion in inventory, delayed new products, and saw net income fall 55 percent in a fiscal year. Then ChatGPT arrived and demand exploded. The lesson is that Nvidia can survive a demand shock if a new wave of computing arrives. But the current cycle is different because so much of the demand is financed with debt.

There are signs that Nvidia is already supporting the neocloud ecosystem. Its cloud service agreements total $26 billion, with $22 billion due by 2031. The company says the spending supports research and its DGX cloud offerings, but analysts have noted that the outlays are large relative to that explanation. One analyst argues that the agreements function as a backstop for neocloud demand. CoreWeave’s second-largest customer in 2024 was Nvidia, which agreed to spend $1.3 billion over four years to rent its own chips. In September, Nvidia signed another $6.3 billion contract with CoreWeave. The timing lines up with a sharp increase in Nvidia’s cloud compute commitments.

If Nvidia is spending more to support its customers, the market may already be shakier than it appears. The commitments are not all on the balance sheet; some are disclosed in notes. At a small scale, such support might be manageable. At $26 billion, it is large enough to affect margins and earnings if treated as a cost of goods sold. The arrangement also creates a circularity that is hard to unwind. Nvidia invests in neoclouds, neoclouds buy chips, Nvidia rents capacity, and the cycle repeats. Each step supports revenue, but each step also increases interdependence.

Competition from Google TPUs and custom chips

The biggest threat to Nvidia’s control may come from its own customers. Google has been building TPUs for years, and its Gemini 3 model was trained on TPUs alone. That achievement rattled the AI industry. Google has signed deals with Anthropic, Salesforce, Midjourney, and Safe Superintelligence, and it is reportedly in talks with Meta. Even the threat of switching to TPUs can help buyers negotiate lower prices. One analysis estimates that OpenAI has saved 30 percent of its total cost of ownership on Nvidia GPUs without deploying TPUs, simply by using the threat of alternatives.

Amazon is building its own chips and is discussing letting OpenAI use them. Microsoft, Meta, and OpenAI are also developing custom silicon. Broadcom is a formidable partner behind some of these efforts. In China, Huawei, ByteDance, and Alibaba are building their own chips. AMD is catching up. By 2027, AMD’s roadmap and Nvidia’s may converge in performance, and AMD is willing to price more aggressively. Nvidia has reportedly made late changes to its 2027 chip design, which suggests it is worried about the competition. A company that feels secure does not usually revise a product that late in the cycle.

Competition could reduce Nvidia’s margins and its ability to subsidize neoclouds. It could also reduce the value of older Nvidia chips, which are less energy-efficient than new alternatives. If Google TPUs and custom chips take share, neoclouds may find it harder to fill data centers with Nvidia hardware. That would weaken the collateral behind GPU-backed loans and make lenders more cautious. The circular financing model works best when Nvidia is the only game in town. It becomes more fragile when customers have credible alternatives.

Neocloud incentives and the annual cadence

Neoclouds and Nvidia do not have perfectly aligned incentives. Neoclouds need to squeeze as much revenue as possible out of each chip before it becomes obsolete. Nvidia needs to sell as many new chips as possible. Nvidia has shortened its product cycle from roughly every two years to every year. That puts pressure on neoclouds to rebuild and upgrade faster, even as their existing assets lose value. The debt load does not shrink when a new generation arrives. The result is a treadmill: borrow, build, buy, depreciate, and borrow again.

Large customers benefit from this arrangement. They can rent capacity from neoclouds instead of building everything themselves, and they can demand the newest chips to keep operating costs low. If the current generation costs half as much to run, why would a customer pay twice as much for older cards? That logic forces neoclouds to keep buying the latest hardware. It also makes it hard for them to stop. The moment they pause, their competitors may win the contracts that justify their debt.

Construction and power constraints complicate the picture. New data centers need purpose-built facilities, cooling, and large amounts of electricity. The industry may need an additional 44 gigawatts of capacity by 2028, but only about 25 gigawatts of new power is expected to come online in that period. Power shortages could extend the life of older chips because new capacity is slow to arrive. But they can also delay projects and strand capital

Source:The Verge News