NVIDIA to launch $500 billion AI infrastructure financing platform

NVIDIA is partnering with six global financial institutions, including BlackRock and Goldman Sachs, to build a $500 billion AI infrastructure platform. The initiative will use "Compute Financing" to provide funds for data centers to purchase GPUs, utilizing the hardware itself as collateral for loans.

NVIDIA is set to build an AI infrastructure financing platform worth up to $500 billion (approximately 675 trillion won) in partnership with six global financial institutions, including BlackRock and Goldman Sachs. The core of this initiative is "Compute Financing," which provides funds to companies looking to build AI data centers to purchase GPUs. In particular, the "GPU-backed loan" method, which uses the GPUs themselves as collateral, is expected to open doors to a new financial market, though the financial sector is signaling cautious verification processes regarding the future value of GPUs.

NVIDIA's confidence and the CoreWeave case: "Even older GPUs can earn money"

GPU-backed loans are not an entirely new concept. Cases already exist in the market where companies holding large quantities of GPUs use them as collateral to raise funds. A representative company is CoreWeave, often referred to as a "Neo-cloud" provider. Building on its experience holding GPUs for cryptocurrency mining, CoreWeave transitioned into a cloud business specialized for AI computing, and in August 2023, it secured a $2.3 billion (approximately 3 trillion won) loan from Blackstone and Magnetar Capital using NVIDIA H100 GPUs as collateral.

NVIDIA shows strong confidence in the long-term revenue-generating capabilities of GPUs. The logic is that while the latest GPUs are essential for training cutting-edge AI models, previous generation GPUs are sufficiently usable during the "inference" stage—applying trained models to services—or the "fine-tuning" stage, where models are modified for specific purposes. NVIDIA explains that it can extend the economic lifespan of equipment by continuously improving its "CUDA" software to increase the efficiency of older hardware.

According to data submitted by CoreWeave to the SEC, a specific scientific research company has extended its contract to use A100-based computing services, which were released in 2020, until 2029. GPU valuation firm Vーカー also estimates the economic lifespan of certain models (GB200 NVL72) to be 9 to 10 years, forecasting that sufficient profit can be generated during this period.

The core question for the financial sector: "Is the income sufficient to repay the debt?"

However, the perspective of financial institutions is cold. Even if NVIDIA emphasizes the utility of GPUs, financial companies must separately evaluate the soundness of individual loans. The questions posed by the financial sector are clear: First, can the revenue earned from the GPUs repay the principal and interest of the loan? Second, if repayment is impossible, can the loan be recovered by selling the GPUs held as collateral?

The mere fact that GPUs can be used for a long time is not enough. Since GPU operations incur massive maintenance costs, such as electricity bills, companies must determine if actual repayment capability remains after deducting these costs from revenue. Accordingly, financial companies focus more on the contract terms with clients rather than the physical lifespan of the GPUs. In an $8.5 billion loan agreement signed by CoreWeave last March, the key basis was not just the GPUs themselves, but the "fixed payment" contract signed with its client, Meta. Thanks to a contract structure where Meta agreed to pay a set amount even if it used less computing capacity, the financial company was able to secure predictable income.

Ultimately, the financial market is judging the safety of loans not only through GPU performance but also through who the customers using those GPUs are and how robust those contracts are.

The gap between uncertain resale prices and the "future of financial engineering"

A larger variable is the "residual value" of the GPUs. In a situation where equipment must be sold to recover loan amounts, no one can guarantee at what price GPUs will trade a few years from now. Unlike existing facilities such as airplanes or ships, the rapid pace of technological advancement in GPUs means there is a lack of historical data to predict future resale prices. Consequently, financial companies are maintaining a conservative stance.

Andrew Chang of the credit rating agency S&P Global Ratings acknowledged that GPUs are used for more than five years but pointed out that the value of the chips must be evaluated very conservatively. Tony Trzcinka of the asset management firm Impax also explained that financial companies typically review loans based on a short depreciation period of 3 to 4 years. Even if the equipment can be used longer, financial companies reduce the value quickly to minimize loss risk, thereby limiting loan amounts or demanding separate guarantees.

Various financial techniques are also being discussed to resolve these uncertainties. An "insurance structure" to cover losses from the disposal of GPUs in the event a Neo-cloud company goes bankrupt is being mentioned, and Goldman Sachs is reviewing a plan to securitize the bonds generated from GPU-backed loans and sell them to other investors. As BlackRock CEO Larry Fink compared this concept to the birth of the mortgage-backed securities (MBS) market in the 1970s and evaluated it as the "future of financial engineering," whether GPUs can establish themselves as massive financial assets beyond simple semiconductors depends on whether their residual value can be proven in the future.

Source: original video (YouTube)