NVIDIA, as a major financier in the global AI construction sector, is facing its first significant stress test due to the tightening of credit markets from key partners.
After developing chips for the early stages of generative artificial intelligence, NVIDIA began to leverage its financial strength to provide funding for AI model developers, data center companies, and a group of so-called neocloud. CEO Jensen Huang's strategy is not only to create the most advanced technology but also to invest in customers and help them obtain loans, thereby enabling NVIDIA's systems to reach more clients.
This means that NVIDIA makes full use of its investment-grade credit rating to provide guarantees for the large-scale projects of OpenAI, enabling the related construction to obtain financing at lower interest rates.
According to NVIDIA's latest quarterly financial report in July, the company provided $36 billion in off-balance-sheet cloud agreements, also known as guarantee arrangements, for neocloud. In addition, it provided guarantees for chips, land, and electricity totaling up to $105 billion for a proposed 10-gigawatt OpenAI data center in Ohio. The company also signed as the lessee of data center leases worth $20 billion in the future, which will then be transferred to third parties.
For the vast majority of these transactions, unless the project participants default, NVIDIA does not need to make actual payments.
The founder and CEO of financial research firm New Constructs, David Trainer, said in an interview: "If the demand for computing power does not meet these insanely high expectations, we will be in trouble."
NVIDIA declined to comment on this matter.
The market environment may soon put NVIDIA's strength to the test. Stocks of companies such as CoreWeave, Nebius, IREN, and neocloud in NVIDIA's portfolio have fallen by at least 30% this month. Among them, the credit default swap spread of the largest company, CoreWeave, has soared to over 11%, forcing investors to pay a higher cost to purchase default insurance.
NVIDIA also supports other neocloud companies that are facing challenges.
The company is an investor in AI, an Australian cloud service provider, Firmus. This month, Firmus suddenly withdrew its IPO plan; NVIDIA has also invested in Nscale in the UK and SB Energy supported by SoftBank. Both companies submitted applications for listing in the United States in September, but there have been no updates on the progress of their plans for several weeks.
NVIDIA closed down on Wall Street for the third consecutive week on Friday, with a cumulative decline of 7.6% so far in October, and is on track to record its worst single-month performance since November last year.
Jensen Huang reiterated his confidence in the market this week. On Thursday, he said regarding CNBC, the demand for neocloud is “insanely high.”
"What's missing in the middle is the working capital that drives this growth," Jensen Huang said to Arjun Kharpal during the GTC conference of the Berlin company, "because demand in multiple fields of AI is very strong, and capital has become scarce."
NVIDIA already holds equity in dozens of companies, including significant investments in OpenAI and Anthropic. As of the second quarter, the value of its investment portfolio amounted to 99 billion US dollars.
Build an ecosystem
Investors told CNBC that a major issue surrounding NVIDIA is how the company can utilize its cash and $5 trillion in market value to get more chips into the hands of customers, as well as what level of risk they are willing to take in this process.
NVIDIA graphics processors are crucial for the training and deployment of the AI model. The company's revenue in its latest fiscal year reached $216 billion, which is approximately eight times higher than when OpenAI launched ChatGPT at the end of 2022.
Before the AI craze, NVIDIA sold graphics cards that could be installed in standard server racks or personal computers. Today, its latest systems are much larger in scale and also more expensive. Morgan Stanley estimates that NVIDIA's flagship product, Vera Rubin, is a 7-foot-high rack containing 72 GPU; it weighs as much as a pickup truck and costs over 7 million US dollars.
Although NVIDIA still dominates the AI chip market, competition is emerging. For example, Google rents out tensor processors ( TPU ) through its vast cloud services, and Supermicro Semiconductor ( AMD ) also launched its first AI rack Helios in July.
Cash is another major advantage for NVIDIA. The company generated $70 billion in free cash flow in the first half of this year, while AMD totaled just over $4 billion; Google's parent company, Alphabet, experienced negative cash flow in the second quarter due to expansion investments.
This gives NVIDIA a huge advantage in building an ecosystem. The company now provides a comprehensive platform that includes software, services, integrated hardware, as well as financial support for many customers.
NVIDIA's latest quarterly financial report for August revealed that the company has off-balance-sheet commitments of over $500 billion in the coming years, with the majority of these obligations maturing before the fiscal year 2029. In the same month, Jensen Huang announced a $500 billion financing deal with six major asset management companies, aiming to treat computing infrastructure as a category of asset.
The agreement promises to provide a residual value guarantee of up to $125 billion for the financing of GPU, in addition to which there is also funding from NVIDIA as a limited partner invested in the private equity fund. The company invested $2 billion in Brookfield's first AI Infrastructure Partners Fund, with the fund aiming to raise a total of $10 billion; furthermore, NVIDIA has also invested an undisclosed amount in KKR's Helix Digital Infrastructure (a $11 billion instrument) as well as in Valor Equity Partners's $5.4 billion dedicated computing power fund.
When announcing the agreement in August, Jensen Huang said about CNBC: “In the past, we manufactured and sold chips, which are technical components that people purchased and used. But now, NVIDIA’s AI factory platform is actually an investable asset.”
Jensen Huang also emphasizes the unique value of his device with a marketing slogan.
In a podcast last week in collaboration with investor Brad Gerstner, he said, “We are the easiest to rent out. We are replaceable, and we are durable.”
20 Years of Hardware in Ohio
While waiting for Wall Street institutions to attract investors and expand the loan pool of GPU, NVIDIA has already been actively promoting record-scale transactions in the infrastructure sector of AI.
NVIDIA’s largest single guarantee is for the facility that OpenAI is about to build in Pike County, Ohio; this project has the potential to become the world’s largest data center. This transaction requires the use of NVIDIA hardware for at least 20 years.
The shift from chips to capital, if OpenAI falls into trouble, will bring significant risks.
If OpenAI fails to make payments or faces insolvency, NVIDIA may take over the lease, seek new tenants, or request that the landlord SB Energy sell the facility; if the recovery or sale price from the latter two arrangements is lower than the minimum price, NVIDIA will provide a guarantee for the difference, up to a maximum of $105 billion. NVIDIA may also pay a one-time sum to SB Energy. On the other hand, if OpenAI provides its own collateral or obtains a sufficiently high credit rating, NVIDIA's guarantee will be terminated.
In the case of CoreWeave, NVIDIA's investment included equity, as well as a commitment to purchase up to $6.3 billion in computing power if no other buyers could be found for CoreWeave.
NVIDIA's guarantees can take various forms, such as ensuring a minimum resale price or a minimum hourly purchase price for the chips when the lender needs to recover assets. For Sharon AI and other neocloud, NVIDIA's transactions also include revenue-sharing agreements.
NVIDIA stated in its quarterly report: "We believe that these AI cloud partners have strong customer demand and solid sales pipelines, but they are limited by the large-scale infrastructure required to meet this demand."
Jensen Huang stated this week that CoreWeave does not require any guarantees from NVIDIA, and moreover, the capital cost of this neocloud is on the decline. The total debt of CoreWeave is approximately $51.6 billion, which includes about $35 billion in notes, bonds, and loans. The yield on its bonds maturing in 2031 and 2032 once exceeded 15%.
Equity Armor Investments Portfolio Manager Brian Stutland stated that the nervousness in the market reflects people's concerns about the stability of this business.
Stutland said that CoreWeave and other neocloud 'will sign future contracts, and everything looks good from the financial reports, but can that contract really remain stable? Could it break down at some point?'
NVIDIA stated that its financial support helps customers build the infrastructure required to meet the strong demand for chips. Additional computing power may also assist the AI laboratories in developing new models and services, thereby further driving the demand for NVIDIA products.
Deepwater Asset Management founder Gene Munster said last month about CNBC: “Their revenue next year will reach 700 billion US dollars. This gives them the ability to provide financing for some of their customers.”
NVIDIA's credit guarantees have mostly only begun to be implemented in the past 6 to 12 months. Their last annual report, released in February, did not mention the AI cloud protocol.
AI, co-founder of infrastructure startup Compute Labs, Warren Hosseinion, stated that NVIDIA does not intend to play the role of an insurance company, "but they are willing to take on that role until the market and industry catch up."
He said, "To some extent, they control the direction of things." The company with Hosseinion has opened up a capital pool for neocloud operators by tokenizing GPU. "They must be very cautious about at what pace they release the technology, because this will largely determine how the market will function around it."
Not a ‘complete panacea’
NVIDIA is actively seeking insurance companies to cover part of the risks associated with potential value losses related to GPU.
Howden, the Executive Director of Product Innovation at the global insurance group, stated: "Insurance companies are not yet the perfect solution to this problem. However, in appropriate circumstances, they are likely to step in and provide assistance, thereby opening up some business opportunities."
According to the research by New Constructs, NVIDIA's off-balance-sheet liabilities had a net present value of over $153 billion in fiscal year 2026, an increase of more than $100 billion from last year.
This type of liability has been increasing throughout the entire AI transactions, with leasing obligations being the main form. Rating agencies such as Moody's have been issuing warnings. The agency stated that in 2026, guarantees will be the fastest-growing type of off-balance-sheet financing arrangement.
Moody's does not include guarantees in the debt rating when assessing corporate buyers, but Moody's Vice President David Gonzales stated that this is "something we continuously monitor."
He said, "These are all complex structures with uniqueness, so there is no one-size-fits-all approach to dealing with them." Gonzales indicates that in the short term, chips are likely to retain a value "higher than the residual contract amount," which will mitigate the risks in many of NVIDIA's financing structures.
However, "the extent to which investment-grade AI companies can withstand such exposures is a question that we assess every day, at all times," said Gonzales.
"That's the core of all our current discussions," he said.
CNBC's Seema Mody contributed to this report.
