Nvidia has announced a sweeping $500 billion financing initiative in partnership with some of Wall Street’s most powerful asset managers—Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR—to transform computer chips into a new investable asset class. The move, which CEO Jensen Huang described as “the first time that technology chips have become an investable asset class,” aims to turn Nvidia’s GPUs into revenue-generating, long-lived, fungible, and flexible assets that can be financed and traded much like mortgages or bonds. But beneath the headline-grabbing numbers lies a complex web of financial engineering, shifting narratives about chip depreciation, and a strategy that some analysts warn echoes the early days of mortgage-backed securities—complete with the same potential for systemic risk.
What is the $500 billion compute financing deal and who is involved?
The deal, still in the memorandum-of-understanding stage, brings together six of the largest financial institutions in the world to provide up to $500 billion in financing for what Nvidia calls “compute.” The consortium includes Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR. These firms will work with Nvidia to create a new asset class based on the computing power of Nvidia’s chips, rather than the physical data centers themselves. Huang has emphasized that “Nvidia compute is not just a chip” because of the accompanying CUDA software, which he claims continuously improves the output of the hardware, extending its economic life well beyond typical depreciation schedules.
However, the deal is not yet finalized. Nvidia previously signed a similar $100 billion memorandum of understanding with OpenAI that ultimately fell through. The current announcement is best understood as a trial balloon—a powerful signal of intent, but far from a binding contract. Even so, the involvement of these heavyweight institutions suggests that the AI financing boom is entering a new phase, one where the underlying assets are no longer just physical infrastructure but the computational capacity itself.
How does ‘compute as an asset class’ differ from traditional GPU-backed loans?
Huang is careful to distinguish “compute” from raw GPU hardware. He argues that Nvidia’s AI factories—complete with accelerated computing, networking, systems software, AI frameworks, and a global developer ecosystem—are more than just chips. But critics point out that the financing model amounts to a familiar concept: the GPU-backed loan. In the Broadcom deal that inspired Nvidia’s announcement, Apollo and Blackstone lent money against a pool of about a million chips, with Broadcom providing a guarantee for senior notes. The arrangement is essentially a secured loan, where the collateral is a set of GPUs. The “compute” label may be a marketing term designed to obscure the fact that the underlying asset—a chip—has a limited lifespan, typically two to five years, far shorter than the 10-year depreciation schedule Huang now promotes.
When asked directly about the return on investment for compute, Huang replied, “The return is in the usefulness of AI,” a vague answer that does little to reassure lenders. In reality, the return for financiers comes from interest payments on the debt, not from any magical appreciation of the chips themselves. The “compute” moniker allows Nvidia to package the same old GPU loans under a more future-friendly brand, one that sounds more like a utility than a rapidly depreciating piece of silicon.
Why did Jensen Huang change his message about chip depreciation?
Last year, Huang told attendees at Nvidia’s AI conference that when the new Blackwell architecture ships in volume, “you couldn’t give Hoppers away.” He dismissed the previous generation as nearly obsolete. Now, he is touting the A100 chip, introduced in 2020, as a “powerful example” of long-lived compute, claiming it “remains in active commercial use” and could have an economic life “toward a decade.” This about-face has caused whiplash among industry observers. The shift is self-serving: by extending the depreciation schedule, Nvidia makes its chips more attractive as loan collateral. Lenders can lend more against an asset that supposedly retains value for 10 years rather than two or three. Short seller Michael Burry has argued that the appropriate depreciation cycle for chips is two to three years; IBM’s Arvind Krishna says five years. Huang’s decade-long claim is at odds with these estimates and with his own previous statements.
Analysts see this as a way to unlock more financing for Nvidia’s customers. If the chips don’t depreciate as quickly as previously thought, then neocloud companies like CoreWeave—which are heavily dependent on Nvidia financing—can borrow more money against their GPU fleets. CoreWeave’s CEO, Michael Intrator, revealed on an earnings call that the company has sold contracts for 2020-era chips extending through 2029, which may be the basis for Huang’s decade claim. But the question remains: can the chips physically and economically last that long, or is this a narrative designed to prop up the financing model?
What are the risks of comparing compute to mortgage-backed securities?
BlackRock CEO Larry Fink told CNBC, “This is the very beginning, like what it was when I started in the mortgage-backed securities market in the 1970s.” The comparison is ominous. Mortgage-backed securities failed when mortgages were overproduced and borrowers defaulted. The AI industry is already showing signs of overbuild: data centers are becoming saturated, and Chinese open-source models require less compute while still delivering strong performance. Both trends threaten the assumption of ever-growing demand for chips. More fundamentally, the frontier AI labs—Anthropic and OpenAI—that are driving much of the current demand are not yet profitable. Anthropic has an annualized revenue run rate of $65 billion, but there is no word on profit. OpenAI and SpaceX are hemorrhaging money. If these companies cannot turn a profit, the demand for compute could suddenly collapse, leaving lenders holding collateral that is worth far less than the loan amount.
Analyst Mark Rubinstein notes that mortgage-backed securities failed when mortgages were overproduced. The same pattern could emerge here: an oversupply of data center capacity and a slowdown in AI adoption could lead to a wave of defaults. The new compute futures, set to be introduced by CME Group in October, may provide a hedging mechanism, but they also create a new layer of speculation. The entire edifice rests on the assumption that AI demand will continue to grow exponentially—an assumption that is far from certain.
How does the deal help Nvidia maintain its competitive advantage?
Nvidia’s move is not just about financing; it is about locking in market dominance. By making it easier for neoclouds to get loans backed by Nvidia chips, the company effectively reduces the cost of financing for its own products relative to competitors like Broadcom, Google’s TPU, or Amazon’s Trainium. As Felix Wang of Hedgeye Risk Management put it, “In effect, they made Nvidia’s product cheaper without really cutting GPU prices.” The consortium also includes a residual value guarantee from Nvidia—up to 25% on some contracts—which reassures lenders but also exposes Nvidia to significant downside risk if the chips lose value faster than expected.
Furthermore, the financing may standardize how Nvidia chips are installed in data centers. Lenders prefer uniformity because it makes it easier to model revenue forecasts and to offload collateral in case of default. Nvidia has begun issuing guidance on ideal settings for its chips, effectively herding data center operators toward a single design. This not only reduces risk for lenders but also strengthens Nvidia’s moat against competitors who offer more customized solutions. The standardization, combined with the financing advantage, creates a powerful flywheel: more loans for Nvidia chips lead to more adoption, which leads to more data from which Nvidia can improve its software, which in turn justifies longer depreciation schedules.
What is the circularity problem and how does this deal address it?
Nvidia has faced accusations of circular financing, where it invests in the neoclouds and AI labs that buy its chips. For example, Nvidia invested in CoreWeave, saved its IPO, and agreed to spend $1.3 billion over four years to rent its own chips from CoreWeave. This arrangement made Nvidia CoreWeave’s second-largest customer in 2024. The new consortium brings in outside capital—Blackstone, Apollo, etc.—to replace Nvidia’s own money in the financing chain. Instead of Nvidia giving CoreWeave a dollar, against which CoreWeave borrows five dollars to buy six dollars of Nvidia chips, now Blackstone gives CoreWeave five dollars directly to buy Nvidia chips. This appears to resolve the circularity at Nvidia’s balance sheet level.
But a deeper circularity remains. The notes issued by these special purpose vehicles are often sold to insurers and pension funds. Half of the consortium members share ownership with the entities that will likely buy the loans, weakening the incentive for rigorous underwriting. As Sascha Steffen of the Frankfurt School of Finance notes, “One layer down, a second circularity has been created.” The ultimate risk is borne by insurers and annuity holders, not by the originators. If the loans go bad, the losses will ripple through the insurance sector, which is already heavily exposed to AI-related debt.
What does the market think about compute as an asset class?
The market’s response has been muted. Jack Ablin, a founding partner at Cresset, a $260 billion family office that invests in Nvidia, noted that compute historically has “the shelf life of lettuce.” Even Stratechery’s Ben Thompson, a frequent tech cheerleader, described Nvidia’s strategy as risky. The launch of compute futures by CME Group in October may provide a mechanism for price discovery, but it also introduces speculative trading that could amplify volatility. The key question remains: will the demand for compute continue to rise, or will it plateau as AI models become more efficient and as the big hyperscalers—Microsoft, Amazon, Google, Meta—build their own data centers with proprietary chips?
SpaceX, a new neocloud player, has announced it will work exclusively with Nvidia chips, after Nvidia took a $21 billion stake in the company. This suggests that Nvidia’s investment strategy is paying off, but it also raises the stakes. If SpaceX or any other major neocloud fails, the losses could be enormous. The resale value of chips in a distressed market is unknown. The big cloud providers are building their own capacity and may not be buyers of second-hand Nvidia hardware. The only backstop is Nvidia’s residual value guarantee, which is limited to 25% of the contract value.
What is the future of the AI financing model?
The Nvidia deal is a bet that AI will follow the trajectory of the internet: a massive infrastructure buildout followed by a wave of profitable applications. But the analogy is imperfect. In the late 1990s, it was clear that streaming video and e-commerce were coming, even if the timing was off. Today, the most concrete AI applications are replacing customer service agents and speeding up coding—useful, but not the kind of transformative, revenue-generating use cases that would justify $500 billion in financing. The AI industry is still subsidized by venture capital and corporate investment. If the model companies cannot achieve profitability, the demand for compute will collapse, and the asset class will prove to be a house of cards.
For now, Jensen Huang is pushing forward with a narrative that is good for Nvidia’s stock price and for keeping the AI party going. The funding consortium gives him a new way to sell chips without needing to cut prices. But the ultimate test will come when the first major loan defaults. If the chips hold their value and the contracts are honored, the model will be validated. If not, the financial system may find itself holding a new type of toxic asset—one that was once hailed as the future of finance. The next few years will reveal whether compute is truly an asset class or just another cleverly packaged debt product.
