Nvidia's Strategy: Does the Math Compute?

Jensen Huang, CEO of Nvidia, is advocating for a paradigm shift, urging a perception of "compute as an asset class" rather than the potentially problematic "GPU-backed loans." This initiative marks a significant foray into financial innovation, with major players aiming to transform the technology landscape.

In a substantial move, financial giants Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR are collaborating with Nvidia to secure an impressive $500 billion in financing. Their collective goal is to solidify "compute" as a recognized asset class within the financial markets.

Huang articulated this vision to CNBC, stating, "This is really the first time that technology chips have become an investable asset class. These are revenue-generating assets now. They’re productive, they’re long-lived, they’re fungible, they’re flexible."

BlackRock CEO Larry Fink echoed this sentiment, drawing a historical parallel: "This is the very beginning, like what it was when I started in the mortgage-backed securities market in the 1970s, and I look upon this as a next future for financial engineering."

However, Huang's current enthusiasm for the longevity of chips presents a stark contrast to his past remarks. Last year, when promoting the new Blackwell GPU architecture, he suggested that once Blackwell shipped in volume, "you couldn’t give Hoppers away," and that "there are circumstances where Hopper is fine. Not many." This abrupt shift to describing chips as "revenue-generating assets" that are "long-lived" has caused considerable surprise.

For the moment, Huang's assertion holds some truth. The cost to rent older chips has been on an upward trend, with Silicon Data projecting continued increases through 2028. An illustrative example saw one cloud service provider nearly double its prices for Nvidia Blackwell B200 chips for a rental customer during a contract renewal.

Fink's comparison to the early mortgage-backed securities market might trigger caution for some. As former hedge fund manager Mark Rubinstein observes, that market faltered due to an overproduction of mortgages. Similar concerns are emerging in the AI sector, with a proliferation of data centers and the rise of powerful yet less compute-intensive Chinese open-source models, both potentially threatening sustained demand for chips. A fundamental question also remains: can leading AI labs like Anthropic and OpenAI, which are fueling much of the current demand, achieve profitability?

It's crucial to note that this initiative is not yet a finalized agreement; it consists of Memorandums of Understanding (MOUs). Nvidia previously signed a $100 billion MOU to invest in OpenAI last year, which ultimately did not materialize. While MOUs allow for significant public announcements, their failure to proceed often goes unremarked. Nevertheless, even as a trial balloon, this development offers intriguing insights into the future of "compute" financing.

So, why the emphasis on "compute"? Huang contends that "Nvidia compute is not just a chip," but rather a comprehensive system that includes its proprietary CUDA software. He states, "That is what makes Nvidia AI factories different" from mere silicon. Huang asserts that "Their value is not fixed at installation: CUDA continuously improves their output; the installed base remains productive well beyond its initial depreciation period."

Yet, chips and software alone do not constitute "compute." They require extensive data center infrastructure, including physical warehouses and power supplies. Huang's definition of Nvidia's system as "a complete AI factory platform including accelerated computing, networking, systems software, AI frameworks and a global developer ecosystem" conspicuously omits these essential brick-and-mortar components.

In this context, "compute" appears to be a re-branding of the familiar GPU-backed loan.

Setting aside the somewhat fantastical notion of an "AI factory," Huang's downplaying of data centers is understandable, given that much of the existing financing has been directed there. Rubinstein highlights Blackstone's $185 billion data center platform, with projections for the long-term ownership market to reach $1 trillion. Huang, however, is primarily concerned with driving sales of Nvidia chips, not the real estate aspect of data centers.

Therefore, "compute" serves as a buzzword for GPU-backed loans. The shift from "GPU" to "compute" might be strategic, as GPUs typically have a shorter lifespan (estimated at two to five years) compared to buildings, and "compute" could also encompass other accelerator-backed loans, such as those for TPUs.

Nvidia's current strategy appears to mirror a $35 billion package Broadcom assembled earlier this summer with Apollo and Blackstone. That deal involved funding "compute" with approximately a million chips as collateral, where Apollo and Blackstone would earn interest, and Broadcom provided guarantees for senior notes. This arrangement was designed to stimulate demand for Broadcom chips, a model Nvidia seems to be adopting for similar reasons.

Consequently, Huang is now enthusiastically championing the extended life of Nvidia chips. As a "powerful example" of how compute can improve over time, he points to the A100 chip, introduced in 2020, which "remains in active commercial use." He further claims that "Customers continue to commit capacity for multi-year deployments, extending A100’s economic life toward a decade." This contrasts sharply with his previous comments about newer chip architectures.

The issue of chip depreciation schedules has long been debated, with no industry consensus. Short seller Michael Burry has suggested a two-to-three-year cycle, while IBM's Arvind Krishna proposes five years. Huang's declaration of a decade-long economic life for his chips is a significant outlier, and it's a statement that could influence lenders.

This directly impacts loan terms, as the borrowing capacity of companies like CoreWeave—a pioneer in GPU-backed loans and an Nvidia client—is tied to the depreciation of its chips. If Huang publicly asserts a 10-year depreciation schedule, banks may be inclined to believe him, which would be highly advantageous for companies seeking loans from this new consortium.

Huang points to rising compute prices, including for the Hopper H100 chip released in 2022. AI industry analyst Brendan Burke confirms these price increases are not an exaggeration, driven by high demand for inference—the process of a trained model analyzing new data. This demand has kept hourly rates for older chips high, and in some instances, even increased them, reversing the expected trend of decreasing prices due to a "major shortage of inference chips."

During CoreWeave's second-quarter earnings call, CEO Michael Intrator mentioned selling GPUs from 2020 architecture under contracts extending to 2029. Given CoreWeave's close ties to Nvidia, this could be the basis for Huang's "decade depreciation cycle" assertion.

This new compute consortium appears to be a highly favorable development for Nvidia.

Further signaling the financialization of compute, CME Group, a derivatives exchange, has announced plans to introduce compute futures in October, pending regulatory approval of the two proposed contracts.

Whether this surge in demand will continue indefinitely is uncertain. The proliferation of data centers suggests that if chips are as fungible as Huang claims, data center providers might face price competition in a saturated market. However, as AI becomes more integrated across industries, a broader range of companies, beyond frontier labs, will require inference capabilities. The pace of this adoption will be critical; a slow uptake could pose challenges for this financing model.

The fundamental question of return on investment for "compute" remains somewhat ambiguous. Huang offers a general answer: "The return is in the usefulness of AI." However, if "compute" in this context primarily refers to GPU-backed loans, then the return on investment would be the standard interest accrued on debt.

The specifics of these contracts are paramount and currently unknown. For instance, in the Broadcom deal that likely inspired Nvidia's announcement, Broadcom only backed the higher-priority senior debt, not all of it. It's plausible that the buyer's contract itself would serve as collateral, with its value varying based on the buyer's financial stability—Microsoft's contracts would be less risky than OpenAI's, given the latter's need for continuous fundraising. Additionally, agreements might include clauses for revenue sharing or other incentives for debt providers. Regardless of these details, this new compute consortium seems highly advantageous for Nvidia.

Vikrant Vig of Stanford University noted last year that most GPU loans used Nvidia chips as collateral, which lowered financing costs for companies due to the chips' liquidity. If these deals finalize, securing financing for Nvidia chips will become even easier. For "neoclouds"—smaller companies that rent out compute, such as CoreWeave, Crusoe, and Lambda—acquiring Nvidia chips would offer a level of support not readily available from competitors.

Felix Wang of Hedgeye Risk Management told Bloomberg that this strategy effectively "made Nvidia’s product cheaper without really cutting GPU prices."

Nvidia has aggressively invested in and provided financing to neoclouds to expand its market reach and reduce the bargaining power of major cloud providers like Microsoft, Amazon, Google, and Meta. A notable example is SpaceX, a significant new neocloud player, which exclusively uses Nvidia chips. It was later revealed that Nvidia holds a $21 billion stake in SpaceX. While SpaceX had been evaluating alternatives, Nvidia's investment may have effectively locked them into its ecosystem for their massive data center buildout.

Brendan Burke points out another interesting consequence of this financing: lenders' preference for loan uniformity could lead to greater standardization in how Nvidia chips are installed in data centers, potentially giving Nvidia a further competitive edge.

GPU performance varies significantly based on setup, making reliable revenue projection difficult for lenders. Nvidia has begun issuing guidance for revenue in "ideal settings," encouraging cloud providers to adopt specific designs. This standardization would simplify lenders' tasks, but also invites scrutiny into customer profitability and the accuracy of Nvidia's models. Burke notes that some forecasts are "very bullish," projecting $70 billion annually per gigawatt, a figure far beyond current industry realities.

Therefore, contract terms might mandate specific settings to enhance the fungibility of data centers equipped with Nvidia chips. This would not only streamline revenue forecasting but also facilitate collateral offloading in the event of a default. Burke suggests that "Most data center operators are highly customized and it’s going to be a major sheep herding exercise to get them to follow one approach," implying that lending conditions could act as a powerful force in standardizing engineering designs.

This strategy also fortifies Nvidia against competition, not just from specialized AI chips like Google's TPU and Amazon's Trainium, but also from CPUs. CPUs can perform inference tasks at approximately one-fifteenth the cost of GPUs, meaning their wider adoption could reduce demand for GPU compute and drive down prices.

Nvidia's pursuit of external capital appears to be a response to accusations of "circular financing," where it acts as a major investor in the very neoclouds and AI labs that purchase its chips. CoreWeave, for instance, has benefited from Nvidia's investment, IPO support, and even a commitment from Nvidia to rent its own chips, making Nvidia CoreWeave's second-largest customer in 2024 with a $1.3 billion rental agreement over four years.

This pattern extends beyond CoreWeave, with Nvidia having broad investments across neocloud companies and committing $30 billion in cloud service agreements, which senior analyst Jay Goldberg of Seaport Research Partners interprets as backstop agreements.

If these new MOUs are formalized, the financing landscape would shift. Instead of Nvidia directly funding a company like CoreWeave, enabling it to leverage that capital for chip purchases, an entity like Blackstone would directly provide the funds for CoreWeave to acquire Nvidia chips.

The consortium includes prominent financial institutions such as Apollo, BlackRock, Blackstone, and Brookfield, among others.