엔비디아는 자체 가속기의 주요 구매처인 AI 연구소들에 약 500억 달러의 투자를 집중하고 있다.
Nvidia has invested nearly $50 billion in artificial intelligence labs that purchase its chips, while helping arrange financing commitments worth more than $500 billion to support the expansion of AI data centres. The model has raised questions about circular financing and Nvidia’s growing financial exposure to its own customers.
Colette Kress, Nvidia’s chief financial officer, told analysts on August 26 that demand from AI labs backed by Nvidia’s balance sheet could account for roughly one-quarter of the company’s business next year. Under this arrangement, Nvidia invests in an AI lab, which then uses the funding—or the credit unlocked by Nvidia’s involvement—to develop data centres equipped with Nvidia chips. Those equipment purchases are recorded as Nvidia revenue, strengthening the company’s cash position and market value and enabling it to make further investments.
Kress said Nvidia has formed partnerships with six major investment firms: Apollo Global Management, BlackRock, Blackstone, Brookfield Asset Management, Goldman Sachs and KKR. The partnerships are intended to establish financing platforms capable of raising more than $500 billion in external capital for AI labs seeking to build data centres and computing infrastructure. Nvidia has also worked with SB Energy to secure land, power and construction capacity for facilities that will exclusively host Nvidia equipment.
The first phase will provide 4.25 gigawatts of capacity for OpenAI. According to Kress, OpenAI’s existing and planned commitments amount to approximately 12 gigawatts of Nvidia computing capacity through 2030. Nvidia will also provide credit support covering nearly 2 gigawatts for a second AI lab whose identity has not been disclosed. However, Nvidia’s financial results statement noted that the partnerships remain subject to definitive agreements. This means binding contracts have not yet been signed and the $500 billion represents a fundraising target rather than capital already secured.
Nvidia is extending a similar form of support to smaller cloud computing operators. The company may commit to renting part of an operator’s computing capacity, providing a guaranteed revenue stream that can help the operator secure financing from lenders. In return, Nvidia receives a share of the revenue generated above the agreed minimum. This allows the company to earn from both the initial equipment sale and the subsequent rental income.
Nvidia does not consider its strategy to be circular financing in the traditional sense. Kress offered three main reasons. First, external lenders evaluate each transaction independently, while Nvidia does not directly provide loans. Second, the company sells its chips to investment-grade customers or businesses supported by financially credible backers. Third, if a customer defaults, the equipment can be transferred to another buyer. Kress said AI labs need this support because their demand for computing capacity exceeds what their current financial resources can fund. Many of these labs are relatively young companies without the long-term contracts or credit ratings that lenders typically require before financing data centres. Their main barrier to growth, according to Nvidia, is not a lack of customers or technology but limited access to computing infrastructure.
The most significant risk arises if one of Nvidia’s backed AI labs fails to meet its financial obligations. In that scenario, Nvidia could lose both the chip sale and the value of its investment. The company argues that the hardware could be reassigned to another customer, although this protection depends on demand continuing to exceed supply. Nvidia says that remains the case in the current market.
During the earnings call, Vivek Arya of BofA Securities asked CEO Jensen Huang about Nvidia funding AI labs that are developing their own processors, including OpenAI’s Jalapeño chip. Huang responded that Nvidia provides a complete computing platform capable of operating across different cloud environments and throughout the full lifecycle of an AI system. Competing chips, he argued, are often designed for a specific service or limited use case. Regarding Nvidia’s investments, Huang said his only regret was that the company had not invested more money at an earlier stage.
Kress told Morgan Stanley analyst Joseph Moore that an AI agent may require between 15 and 100 times more computing power than a person using the same system. Huang also said he believes AI usage shifted to become predominantly agentic during the past month, although Nvidia did not provide data supporting that assessment.
Based on its demand expectations, Nvidia forecast revenue of $108 billion for the current quarter. The company also issued a preliminary projection of approximately 70% growth for the financial year ending in January 2028. Kress said Nvidia’s growth is currently constrained by supply capacity rather than customer demand. Kress separately warned that memory prices are increasing faster than Nvidia had expected, partly because of the rapid expansion of AI infrastructure. Nvidia expects its gross margin to decline to 74% in the current quarter before reaching a low of between 71% and 72% in the fourth quarter. The company is scheduled to report its next financial results on November 17.