Wall Street is betting that AI chips could overturn one of finance's basic rules that "rapidly advancing technology rapidly loses its value."
The Financial Times (FT) evaluated this way Nvidia's $500 billion (about 706 trillion won) AI infrastructure financing platform that it is pursuing with six financial firms—Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR. What makes this plan, which includes major Wall Street investment banks (IBs) and asset managers, stand out is the idea of issuing loans using graphics processing units (GPUs), which quickly become obsolete as technology advances, as collateral.
GPUs are core components of AI data centers, but the prevailing view in finance until now has been that the rapid generational turnover driven by technological progress makes them unusable as collateral for long-term loans. Nvidia and Wall Street aim to redefine GPUs as assets that generate cash over long periods, like power plants or aircraft, and open a new market for collateralized lending.
This shift in thinking was possible because of Nvidia's software ecosystem, CUDA. According to the FT, a senior Wall Street executive noted that based on CUDA's software compatibility and continued support, Nvidia's older GPUs such as the A100 and H100 can be redeployed for new AI computation and continue to generate revenue. To reduce the risk of declining GPU collateral value from technological progress and generational shifts, Nvidia is considering providing residual value support of up to 25% of each project's investment amount. The idea is to make it easier for financial firms to lend by covering part of the loss if GPU prices fall more than expected at loan maturity.
After cash and corporate bonds comes "GPU collateralized lending"
The "GPU finance alliance" between Nvidia and Wall Street means the center of AI investment funding is moving in earnest from internal cash at corporations to the capital markets. In the early stages of AI investment, hyperscalers such as Microsoft (MS), Google, Amazon, and Meta, which generate massive operating profits, built data centers with cash on hand. More recently, as investment volumes have begun to exceed their own cash generation, corporate bond issuance has surged. But concerns about debt burdens from large borrowings have heightened financial market anxiety and raised questions about the sustainability of AI investment. This is the backdrop for a model that seeks to draw new funding sources such as insurers and pension funds by using GPUs as collateral assets.
The basic structure of GPU finance resembles real estate collateralized lending. When infrastructure operators such as neoclouds, AI corporations, and hyperscalers purchase Nvidia GPUs and build data centers, financial firms lend money by taking both the GPUs and the clients' long-term use contracts as collateral. Principal and interest are repaid with GPU leasing fees and usage fees. If the borrower falls into default, the financial firm can repossess the GPUs and sell or lease them to another operator. The FT projected, "Some financial firms could set up semiconductor finance specialists and structure GPU finance in a way similar to collateralized loan obligations (CLOs), which bundle leveraged loans used for private equity buyouts." If GPU-related loans are tranched into multiple tiers by credit risk, investors can invest according to their risk appetite and expected return.
Some expect the emergence of structured finance in which a special purpose vehicle (SPV) owns Nvidia GPUs and data centers and issues asset-backed securities based on the resulting lease and usage fees to attract long-term funds from insurers and pension funds. Up to now, fledgling AI corporations and neoclouds have relied on high-rate private credit, but going forward they could access the capital markets more broadly and cheaply through products that securitize GPU collateralized loans.
Nvidia's AI dominance could grow stronger
However, GPU finance carries the risk that a significant portion of collateral value depends on Nvidia's CUDA platform. Financial firms assign higher collateral value to Nvidia GPUs, which have a large secondary market and guaranteed CUDA compatibility. Operators that secure funds at lower rates then choose Nvidia GPUs again. The advantage of the technology ecosystem converts into a gap in financing costs, and the lower financing costs in turn reinforce Nvidia's technological monopoly, completing a circular structure.
That said, in this structure, if Nvidia's technological dominance weakens, the stability of GPU collateralized lending is also threatened. If alternatives such as Broadcom's application-specific integrated circuits (ASICs) and Google's tensor processing units (TPUs) spread, and if open software ecosystems grow and reduce developers' dependence on CUDA, the profitability and residual value of older Nvidia GPUs fall. If demand for AI computation slows, GPU rental income and collateral value decline together, potentially weakening borrowers' repayment capacity and financial firms' recovery prospects.
Will GPU structured finance inflate an AI bubble?
Above all, GPU finance could become a channel that amplifies an AI bubble. If GPU collateralized loans are securitized into multiple credit ratings like CLOs, more investors can be drawn into the AI infrastructure market, but as risk is spread out, the true scale of distress and the ultimate risk bearers can become unclear. It is reminiscent of the period before the 2008 global financial crisis, when optimism that home prices would keep rising underpinned the securitization of subprime mortgages into mortgage-backed securities (MBS) and collateralized debt obligations (CDOs), fueling credit expansion and a housing bubble. GPU collateralized loans differ from home mortgages in assets, borrowers, and product structure, so they cannot be equated with the subprime crisis, but it is similar in that optimistic outlooks can expand both credit supply and capital investment, and that risk can spread across the financial market through structured products.
If Nvidia's GPU finance succeeds, AI Semiconductor will overcome short product cycles and take root as infrastructure assets. OpenAI and new cloud firms will lower funding expenses, and Nvidia will secure a larger market and stronger monopoly power. Conversely, if demand for AI computation or CUDA's dominance falters, Nvidia's technological risk will immediately transfer to Wall Street's credit risk.
Will Nvidia's technological dominance last longer than the maturities of GPU collateralized loans? The outcome of Wall Street's bet will shape the AI investment boom and the course of the global economy.