To facilitate the financing of artificial intelligence data centers, Nvidia is working on a model in which major financial institutions such as Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR can provide capital of up to $500 billion in total. While the announced amount alone is remarkable, the longer-term outcome of the plan for the tech sector could be the creation of a strong second-hand market for Nvidia GPUs. The company undertakes to cover a certain portion of the losses of the financing institutions if the graphics processors used as collateral in data center investments cannot maintain their expected values. Thus, Nvidia aims not only to increase the sales of its new generation accelerators, but also to create a broader market where used hardware can maintain its economic value for several years. If this approach is successful, legacy Nvidia GPUs could become more accessible options for smaller cloud providers, companies, researchers and startups.
The warranty that Nvidia will undertake under the plan is not unlimited. If the data center operator cannot fulfill its loan obligations and the creditor is forced to sell the GPUs pledged as collateral, Nvidia will be able to cover up to 25 percent of the difference arising from the hardware being below the projected value in the books. For financial institutions, this mechanism reduces the risk of accepting artificial intelligence hardware as collateral, the value of which is uncertain after a few years. Nvidia, on the other hand, is trying to attract more corporate capital to data center projects without revealing its entire balance sheet. However, if the demand for AI processing capacity declines significantly, a drop in GPU prices could increase Nvidia’s warranty obligations while simultaneously keeping its revenue from new hardware under pressure.
Nvidia wants to make the second-hand GPU market part of its artificial intelligence infrastructure
In the financial world, this corresponds to a type of risk called “wrong-way risk”. The main problem is that Nvidia’s payment obligation can grow under conditions where the company’s core business also weakens. If AI investments slow down faster than expected or more efficient technologies reduce the need for existing GPU infrastructure, second-hand hardware values may fall below estimates. Despite this, Nvidia CEO Jensen Huang argues that the company’s processors should not be considered as rapidly depreciating products like classic PC hardware. Because AI servers can be transferred to different customers, cloud services or operators, they can continue to generate revenue throughout their lifetime, and this large pool of users can support the residual value of the hardware, according to Huang.
The model brought with it comments comparing Nvidia to the financing methods used by Lucent Technologies in the past. Telecommunications equipment manufacturer Lucent provided financing to its customers to purchase its own products during the dotcom bubble and encountered serious problems when the market crashed. Nvidia’s situation is not entirely independent of this example; The company has supported artificial intelligence laboratories such as OpenAI and Anthropic, as well as infrastructure providers such as CoreWeave, Nebius, Firmus and Lambda, with billions of dollars of investments and financing connections. According to Bloomberg’s calculations, the total size of the interconnected deals the company is working on in the summer reaches up to $750 billion. However, in the new model, most of the capital and underlying financing risk is undertaken by independent institutional investors rather than Nvidia.
In his statement on X, Huang directly responded to criticisms about whether the regulation in question was “cyclical financing”. According to Nvidia’s CEO, the goal is to attract long-term independent institutional capital into the AI infrastructure market. The fact that the company protects only a certain part of the GPU value is one of the main factors that distinguishes it from the Lucent example. However, the model is based on the assumption that there will be sufficient demand for Nvidia’s hardware in the future. The validity of this assumption will depend on the continued growth in the use of AI services and the ability to run older generation accelerators economically after new processors are released.
The need for new financing sources is closely related to the high infrastructure expenditures of large technology companies in recent years. While companies like Oracle have increased their debt loads, Google has resorted to new share issues, and Meta is using large amounts of cash for data center investments. Microsoft CEO Satya Nadella’s suggestion of a book addressing the financial crisis following the railway investments in 1873 in his last financial results meeting also coincides with the discussions about the scale of today’s capital expenditures. As in the railway era, the possibility of rapidly growing infrastructure overestimating future demand also applies to the artificial intelligence sector. Therefore, the success of Nvidia’s plan depends not only on financial engineering, but also on the actual use of the processing capacity installed in its data centers in the coming years.
For Nvidia, the strengthening of the second-hand GPU market may partially offset this uncertainty. Companies, research institutions, and startups that don’t need the latest Blackwell or later architectures can perform inference, model fitting, or smaller-scale training tasks using older GPUs at low cost. The widespread use of vulnerability-based artificial intelligence models makes it no longer necessary to use the most expensive and newest hardware in every workload. On the other hand, factors such as energy consumption per performance, memory capacity and software support will continue to determine the economic life of old accelerators. If the efficiency gap of new GPU generations grows too quickly, the electricity and operating cost of running older hardware could negate the advantage despite the lower purchase price.
Nvidia’s plan, which could reach up to $500 billion, cannot therefore be considered merely as a financing arrangement that will enable the construction of more data centers. The company also wants to create a more liquid hardware market where GPUs can be bought and sold throughout their lifetime and transferred to different user groups. This has a distinct advantage for Nvidia; A GPU whose resale value remains strong can also become a more attractive asset for data center operators during the initial purchase. By contrast, a sharp decline in AI demand could simultaneously impact both new GPU sales and the resale values that Nvidia has pledged to protect. This is exactly where the balance of the plan is established: While Nvidia can attract more corporate capital to its artificial intelligence infrastructure, in return it financially guarantees that its aged GPUs will have a real and sustainable use in the future.
Join Channel