The "white glove" game of Neocloud: NVIDIA is selling chips, while they are taking out usurious loans.

The better Nvidia's AI chips sell, the more debt a specific category of companies accumulates.

CoreWeave is the most prominent player in this group. As of the end of June 2026, its total consolidated debt and lease obligations stood at approximately 52 billion U.S. dollars; according to market consensus forecasts, the cumulative negative free cash flow from 2026 to 2028 may reach 71 billion U.S. dollars. Its model of "borrowing at high interest rates and earning thin returns on assets" was directly called the "subprime mortgage" of the AI sector by analysts at DA Davidson.

But CoreWeave is not an isolated case. A group of companies known as "Neocloud" including Nebius, Lambda, and Together AI are all doing the same thing: buying GPUs from Nvidia, leasing them to Microsoft, OpenAI, and Meta, and then taking on huge amounts of debt. Gartner predicts that by 2030, more than 80% of Neocloud service providers will be forced to cease independent operations and exit the mainstream market.

To understand this track, there is no need to look at the computing power parameters of GPUs. All you need is a balance sheet, and a debt noose that is constantly tightening.

White Glove: Helping "Investor Daddies" Expand Their Balance Sheets

The essence of Neocloud is a "white glove".

Nvidia needs to sell chips, but the procurement pace of hyperscale cloud vendors such as Microsoft, Amazon, and Google is constrained by their own capital expenditure, and they cannot accept orders infinitely. Thus, Neocloud emerged: they borrow high-interest debt from banks and private equity markets, use the money to buy Nvidia's GPUs, and then lease the computing power to customers in need.

In this structure, Nvidia's role has long gone beyond that of a simple supplier. It is more like a "financial market maker" that provides implicit guarantees. Taking CoreWeave as an example, Nvidia holds approximately 11% of its equity, sells GPUs to it, and promises to purchase up to 6.3 billion U.S. dollars of remaining cloud computing capacity before 2032 — which means that Nvidia will underwrite the unsold computing power of CoreWeave.

Around July 2026, Nvidia further launched the "GPU lease income underwriting" and revenue sharing model for Neocloud, providing implicit endorsement for the financing of these companies. This model is called "circular financing": Nvidia invests in or guarantees Neocloud, Neocloud uses this credit endorsement to borrow money with dozens of times of leverage from banks, and then places crazy orders with Nvidia.

Chip sales are pushed up, Nvidia's stock price is pushed up, while leverage and default risks remain on Neocloud's balance sheet. An analyst summed it up very directly: Nvidia did not lend directly to Neocloud, but "makes money on both sides" through income guarantees and revenue sharing. The profits of chips go to Nvidia, the interest on debt goes to banks, and the risks go to Neocloud's bondholders.

Structural Breakage: Borrowing Short to Lend Long, Borrowing Expensively and Leasing Cheaply

The reason why this white glove is fragile is not whether there is demand for AI, but that its financial structure has been inverted from day one.

On the cost side, there is short-term, high-interest, and constantly rolling debt. A 2.6 billion U.S. dollar delayed-draw term loan completed by CoreWeave in August 2026 has an interest rate of Term SOFR plus 5.50%; in contrast, a 775 million U.S. dollar secured loan of Nebius in the same period has an interest rate of only SOFR plus 2.50%. The more severe problem lies in the existing bonds: a bond maturing in 2031 has a coupon rate of 9%, and its secondary market price has fallen to 79.9 cents on the dollar, corresponding to a yield of 15.6%; another bond with the same maturity has a yield of 16.2%.

On the revenue side, there are long-term, price-locked, low-margin computing power lease contracts. CoreWeave's second-quarter revenue increased by 112% year-on-year to 2.6 billion U.S. dollars, and long-term committed contracts contributed the vast majority of the revenue, but its capital expenditure reached 9.4 billion U.S. dollars, more than three times the revenue of the quarter, with a net loss of 626 million U.S. dollars, of which the interest expense on GPU-backed debt alone amounted to 640 million U.S. dollars.

What is more fatal is the depreciation risk on the asset side. GPUs are not factories, and their technology iteration speed is extremely fast. A top-tier chip purchased for hundreds of thousands of dollars today may be significantly devalued by a new generation of products two years later. The long-term contracts signed by Neocloud often last for three to five years, but the actual economic life of the underlying hardware assets may only be two to three years. Once the launch of new-generation chips leads to a drop in rental rates, these long-term contracts will turn into a serious financial burden.

The problem with this structure is not how high the interest rate of a certain debt is, but that the maturities of assets and liabilities are completely mismatched. Neocloud uses short-term debt that matures intensively within two to three years to support a series of price-locked long-term computing power contracts. Since 2026, the overall occupancy rate of general-purpose computing power cabinets in China has been under pressure, and the vacancy rate in some regions is on the high side; the lease price of Nvidia H200 has also experienced a short-term decline. If rental income cannot cover interest, the debt rollover chain will face breakage.

Silicon Valley Bank-style Liquidity Crisis

This mismatch of "borrowing short to lend long, borrowing expensively and leasing cheaply" has an extremely precise analogy in financial history: Silicon Valley Bank in March 2023.

Silicon Valley Bank took in short-term, withdrawable-on-demand deposits and held long-term bonds whose prices were sensitive to interest rates. When interest rates rose and deposits flowed out, it was forced to sell assets at a loss, quickly fell into liquidity depletion and declared bankruptcy. The structure of Neocloud is exactly the same: on the liability side, there are short-term, floating-rate debts that require constant refinancing; on the asset side, there are long-term, fixed-price computing power contracts and rapidly depreciating GPUs. Once the credit market tightens, or customer demand expectations reverse, the refinancing window closes, and the company will quickly fall into a systemic solvency crisis under the squeeze of cash flow.

In order to cover this cash flow pressure, some Neocloud players have room for the operation of "interest capitalization" in their financial statements, which counts a large amount of interest expenses during the construction period as capital expenditures, thus delaying the manifestation of losses on the income statement. But capitalization can only postpone the pain, not eliminate the debt. When these projects enter the operation period, the high interest will directly hit the income statement and completely break through the meager rental spread.

The CDS market of CoreWeave has already given its pricing. On October 14, 2026, its 5-year CDS spread once hit 1194.5 basis points, implying a default probability of more than 60% within five years. The three co-founders cashed out a total of 3.3 billion U.S. dollars in the same time window, and nearly 40% of one of their personal wealth has been converted into cash and equivalents. The founders chose to monetize their equity, while bondholders stayed in place to bear the default risk — the two actions point to the same direction: risks are settling to the lowest layer of the balance sheet.

Customer Concentration Dependence and Survival Dilemma of Second-tier Players

In addition to the fragility of the financial structure, Neocloud also faces two major mountains in its business logic: high customer concentration and homogeneous competition.

Taking CoreWeave as an example, most of its revenue has long been dependent on giants such as Microsoft, OpenAI, and Meta. This extremely concentrated customer structure means that Neocloud has no initiative in bargaining. Microsoft is not only CoreWeave's customer, but also its potential competitor. With the gradual maturity of Microsoft and Amazon's self-developed AI chips (such as Maia and Trainium), and large manufacturers starting to adjust their computing power procurement strategies, the sustainability of Neocloud's orders is facing severe tests. Once large customers reduce the proportion of external procurement, Neocloud's long-term contract revenue will be cut off instantly.

Those second-tier Neocloud companies that did not get Nvidia's special "underwriting" commitment (such as Lambda, Together AI, etc.) face a much harsher living environment. They face higher borrowing interest rates, shorter payment terms, and stricter collateral conditions. At a time when the supply of computing power is gradually becoming oversupplied, in order to seize the remaining small and medium-sized customers, they have to launch a price war, further compressing the already meager profit margin. Many second-tier companies are actually using the cash on their books to resist depreciation, waiting for the next round of financing to extend their lifeline.

This Is Not an AI Bubble, But a Bubble of Financing Structure

A distinction needs to be made: the plight of Neocloud does not mean that AI computing power demand has disappeared.

CoreWeave's backlog of orders still exceeds 100 billion U.S. dollars, and long-term contract revenue accounts for a relatively high proportion, so the demand is real. The problem lies in the way to meet the demand. Microsoft and Meta can use operating cash flow (hundreds of billions of dollars in net profit every year) to build data centers, and Apple does not even need to participate in this heavy-asset competition. But Neocloud must use short-term high-interest debt to catch up with a long-term infrastructure cycle, taking all the maturity risk and interest rate risk on its own shoulders.

Nvidia gets chip revenue and ecological control in this game, banks get high-interest interest, and large customers get computing power. The only one with no way out is the white glove itself.

The construction of AI infrastructure takes time, but the maturity date of debt is rigid. When the financing cost exceeds the affordability of computing power pricing, Neocloud will not decline slowly. It will be like Silicon Valley Bank, on a quiet morning, finding that it cannot borrow the next sum of money, while the GPUs in hand are still depreciating rapidly. This prosperity built on extremely high leverage can only be digested by real and continuously growing AI commercial profits in the end, otherwise, the bell of debt will ring sooner or later.

This article is from the WeChat Official Account "Tech Business", author: Tech Business, authorized for release by 36Kr.