SemiAnalysis: NVIDIA's 'Backstop' Plan Could Drive a $7 Trillion AI Debt Market

Original author: Zhao Ying

Source: Wall Street Journal

On July 6, the renowned semiconductor research firm SemiAnalysis posted six consecutive tweets on X, revealing that NVIDIA's Kyber NVL144 rack has been delayed by over 12 months due to manufacturing challenges with its intermediate PCB layer. Asian AI hardware supply chains responded with sharp declines.

NVIDIA subsequently responded that "the roadmap remains unchanged," but did not disclose specific progress details.

The controversy is far from over. On July 7, SemiAnalysis released another paid long-form article targeting NVIDIA—but this time, he is no longer playing the role of a bear.

SemiAnalysis estimates: By 2029, global AI debt financing will exceed $7 trillion. What does $7 trillion mean? It will be second only to the U.S. mortgage market (approximately $13 trillion), becoming the world’s second-largest asset-backed debt market.

What is NVIDIA's role? SemiAnalysis has revealed NVIDIA's strategic initiative called the "Backstop" program. NVIDIA is using its AA/Aa2 investment-grade credit rating to provide minimum revenue guarantees to AI computing rental providers, thereby enabling banks to extend loans. In other words, NVIDIA is acting as the ultimate lender and insurer for the entire AI ecosystem, booking large volumes of sales while absorbing part of the risk associated with insufficient downstream demand. SemiAnalysis has compared NVIDIA to the "central bank of AI."

Regarding the discussion on Platform X about whether SemiAnalysis is “bearish on NVIDIA.” The firm said:

I have not published any positive or negative views on NVIDIA's stock; I merely accurately captured supply chain and technical details, leaving market participants to trade as they see fit.

SemiAnalysis believes that AI infrastructure development is creating a credit market worth trillions of dollars. By 2029, AI-related outstanding debt could reach approximately $7.1 trillion, surpassing all other U.S. asset-backed debt markets except for the U.S. mortgage finance market.

This debt primarily stems from two categories of capital expenditures. One is AI IT capital expenditures, including GPUs, networking, storage, and accompanying CPUs; the other is AI data center capital expenditures, including the infrastructure required to house these GPUs, such as data centers, power, and cooling systems.

In the past, cloud giants like Google, Amazon, Meta, Microsoft, and Oracle primarily relied on their own cash flow to build AI clusters. But over the past year, Oracle, Meta, and even Google have begun to rely more heavily on debt. As project scales continue to expand, market constraints are no longer just about securing GPUs or simply finding data center space—they’re about accessing sufficiently cheap and long-term financing.

SemiAnalysis concludes that the financing model for AI capital expenditures is changing. Cloud giants' balance sheets are not infinite, and if all AI clusters rely on backing from a handful of investment-grade cloud providers, new projects will eventually hit a credit bottleneck.

SemiAnalysis breaks down AI project financing into a "trinity": capital, underwriting contracts, and data centers.

First is capital. Lenders typically require long-term take-or-pay contracts with investment-grade cloud providers or similar credit guarantees before they are willing to extend loans. In other words, what lenders truly care about is not Neocloud’s own creditworthiness, but the creditworthiness of its underlying clients.

Second is underwriting. Neocloud often needs to prove it can pay the GPU deposit and secure devices in order to acquire customers. But to obtain equity funding, it must first demonstrate that it has customers and loans. This creates a circular dependency that easily traps the project in its early stages.

Third is the data center. Neocloud either secures customer contracts and financing to persuade data center operators to lease capacity, or builds its own data center. The latter places greater financial pressure and requires a longer timeline.

This model locks the market into a template of “five-year contracts backed by cloud giants.” The problem is that many VC-backed AI startups and inference providers need short-cycle, large-scale compute—not five-year long-term contracts. Inference providers, in particular, are reluctant to take on long-term price and demand risks, and in many cases would rather forgo compute than sign leases longer than one year.

NVIDIA has launched the "Backstop Program" to address this funding gap.

According to SemiAnalysis, NVIDIA has guaranteed GPU rental income to Neocloud. If third-party customer demand is insufficient, NVIDIA commits to purchasing computing power at a predetermined price; if Neocloud rents out computing power at a higher price, NVIDIA shares a portion of the excess revenue.

These arrangements typically have a six-year term and provide a minimum revenue guarantee for underlying GPU capacity based on a pre-agreed price curve. Neocloud can still lease computing power to any customer and offer more flexible lease terms. NVIDIA’s guarantee is only triggered if market demand is insufficient to rent out the capacity at market prices.

This is the origin of the metaphor of the "AI Central Bank." NVIDIA does not issue currency; rather, it acts as a de facto buyer of last resort and credit guarantor within the AI computing credit system. Lenders can assess the worst-case scenario of projects based on NVIDIA’s AA/Aa2 credit rating, making them more willing to extend loans.

For NVIDIA, this helps expand the base of GPU buyers. If demand for GPUs relies solely on a few hyperscale cloud providers signing five-year take-or-pay contracts, GPU demand will quickly hit financing constraints; meanwhile, these cloud providers are still hedging against NVIDIA’s systems with their own custom chips. Supporting Neocloud and more enterprise customers effectively opens new financing channels for GPU demand.

SemiAnalysis emphasizes that Neocloud is not using NVIDIA credits for free. Under the downside protection structure, Neocloud sacrifices a portion of its upside gains in exchange for project financiability.

In an example price curve, the 6-year average floor price is $2.36 per GPU per hour. Assuming the first-year, one-year lease price for the GB300 is $6.75 per hour and the first-year floor price is $3.68 per hour, the difference between the customer price and the floor price is $3.07. If NVIDIA takes 40% of the amount above the floor price, NVIDIA receives $1.23, and Neocloud receives $1.84, resulting in Neocloud’s actual first-year revenue of $5.52 per hour, which is lower than the $6.75 per hour without a floor price.

Over six years, NVIDIA has averaged a take rate of approximately 18% in this scenario. The project IRR for Neocloud would also decline. With NVIDIA’s guarantee and primarily one-year short-term leases, the project IRR is 25.4%; without guarantee but with successful financing and leasing, the IRR could reach 40.7%.

The key lies in the worst-case scenario. If demand is insufficient, Neocloud can only lease its computing power to NVIDIA, potentially resulting in project returns close to zero or even slightly negative. Lenders do not require the project to profit in the worst case—only that it can still repay its debts. This is precisely why the viability of the debt ultimately hinges more and more on the reliability of NVIDIA’s backstop.

This is the core issue investors should focus on: NVIDIA’s arrangement may briefly boost GPU sales and Neocloud expansion, but if compute demand falls short of expectations, NVIDIA will absorb the revenue gap. The debt may not appear directly on NVIDIA’s balance sheet, but the safety margin of the financing model is increasingly concentrated in NVIDIA’s credit.

SemiAnalysis says that currently, pricing in the GPU financing market is not primarily based on Neocloud’s own creditworthiness, but rather on who has signed long-term off-take agreements.

CoreWeave serves as a reference. Its five-year unsecured bond yield is approximately 10%; however, in the $8.5 billion DDTL 4.0 delayed-draw term loan backed by Meta, the fixed-rate portion costs about 5.9%, only 90 basis points above Meta’s five-year bond yield of approximately 5.0%. This 90-basis-point spread roughly reflects the market’s pricing of CoreWeave’s execution risk.

If Neocloud exits long-term cloud provider take-or-pay agreements, its financing costs will rise significantly. For leading Neocloud providers, unsecured financing may require an interest rate of approximately 10%, about 4 percentage points higher than backed financing. At a loan-to-value ratio of 70% to 80%, financing costs would increase from 5.62% to 10%, reducing pre-tax profit margins from 14.8% to 5.4%.

NVIDIA’s guarantee will position pricing between the total yield of approximately 5.9% backed by current cloud providers and the approximately 10% yield of CoreWeave’s unsecured bonds. Banks prioritize the debt service coverage ratio (DSCR). For projects with NVIDIA’s guarantee, loan sizes are typically calculated under scenarios where the guarantee is triggered, requiring a DSCR of at least 1.3x in the early years, corresponding to a loan-to-value ratio of generally 70% to 80%.

The publicly disclosed NVIDIA-backed projects are currently concentrated in the Asia-Pacific region.

The first is SharonAI’s 72MW AI facility in Australia. Announced in June 2026, the project plans to scale up to a maximum of 40,000 GB300 GPUs under a six-year take-or-pay agreement. SharonAI disclosed a total guaranteed value of $4.88 billion, equivalent to an average six-year base price of approximately $2.33 per GPU per hour.

Another is Firmus’s 360MW AI cluster in Batam, Indonesia, possibly located at DayOne’s facility in the Kabil Industrial Tech Park. Announced on June 29, 2026, this project indicates NVIDIA’s backing is scaling up significantly.

Firmus expects the project to generate $25 billion to $30 billion in customer revenue over six years, targeting AI-native companies, enterprise clients, and inference service providers, with flexible lease terms. However, before deploying GPUs, Firmus still needs to determine whether to partner with a data center provider or continue building its own.

SemiAnalysis also noted that NVIDIA is not the only GPU manufacturer to use buyback arrangements. Last year, AMD offered similar arrangements to customers such as AWS, OCI, DigitalOcean, Vultr, Tensorwave, and Crusoe: if Neocloud was unable to fully sell its capacity, AMD was willing to lease back a portion of the purchased AMD GPUs under long-term contracts for internal software development.

At the time of this article’s publication, SemiAnalysis is also facing controversy.

On the morning of July 6, SemiAnalysis posted a series of tweets on X claiming that NVIDIA’s Kyber NVL144 rack architecture faced significant delays, pushing its timeline back by over 12 months to 2028. The news drew attention ahead of market open and caused declines in multiple AI hardware supply chain stocks in Japan, South Korea, and Taiwan. NVIDIA subsequently responded, stating that its product roadmap remains unchanged and denying any impact on core progress.

This made it easier for the market to interpret SemiAnalysis’s follow-up article as either bearish or bullish on NVIDIA. In response, SemiAnalysis stated on X that it had not published any positive or negative views on NVIDIA’s stock, but had only shared details about the company’s supply chain and technology.

Crackerjack Finance countered the "bearish" interpretation, stating that the SemiAnalysis chart shows actual data for the second half of the year was 20% higher than market expectations, leading to an estimated EPS of approximately $15 next year and a stock price range of $300 to $400. THE Grand Poobah commented, "Three-way circular financing seems no longer sufficient," pointing to market concerns over the increasing complexity of financing structures.

The issue is that AI-related assets have experienced years of price appreciation, with valuations and expectations at elevated levels. Any signal of supply chain risk or changes in financing structures will be rapidly amplified. SemiAnalysis’s clarification that it did not directly provide stock recommendations will coexist with ongoing debates about the market influence and credibility of its supply chain leaks following the Kyber NVL144 incident.

For investors, the true meaning of this "long read" is that the AI competition is no longer just about "who has GPUs," but about "who can piece together GPUs, debt, customer contracts, and data centers all at once." NVIDIA's backstop mechanism may continue to amplify GPU demand, or it may concentrate more of the tail-end pressure of the AI debt cycle onto NVIDIA's own creditworthiness.