NVIDIA's Next Big Move: Partnering with Wall Street on a $500 Billion Bet—Can It Escape the Fate of GPU Depreciation?

Author: Jae, PANews

Tech hardware has always been subject to rapid depreciation, but AI chips are attempting to break this financial iron law: "the faster the iteration, the faster the devaluation." GPUs are beginning to exhibit characteristics of financial assets; they can generate continuous rental income, be used as collateral for financing, and have their residual values calculated based on future cash flows.

On August 10, NVIDIA joined forces with six major Wall Street institutions, including Apollo, BlackRock, and Blackstone, to plan an AI infrastructure financing platform with a scale of up to $500 billion, moving the logic of "computing power assetization" from concept to reality.

In fact, over the past six months, GPU rental prices in the non-hyperscale cloud (Neo-Cloud) market have collectively bottomed out and rebounded, repairing the capital market's discounted expectations for the long-term value of computing power.

Collateralized financing, rental curves, and secondary market residual values are forming an interlinked closed loop: GPUs enter the financial system as collateral, computing power rentals generate cash flow, and future rents are used to repay debt. If this cycle proves viable, GPUs could become a new class of standardized assets on Wall Street.

The prerequisite for the assetization of computing power is stable and predictable rental income.The Neo-Cloud market price index released by Silicon Data is sending clear signals of a recovery.

Over the past six months, hourly rental rates for NVIDIA's mainstream GPUs across various generations have shown a strong trend of bottoming out and recovering. Specifically, the H100, the workhorse for model training and advanced inference, has seen its hourly rental rate rebound from a low of around $2 at the end of last year to $2.72. The H200, equipped with larger VRAM and focused on long-context inference, commands an hourly rate of $3.29. The Blackwell architecture's B200, just entering its shipment phase, maintains a high level of $5.61 due to initial capacity scarcity. Even the A100, released six years ago, has maintained a steady hourly rate of $1.65 thanks to consistent demand in inference and fine-tuning scenarios.

B3Labs, a venture affiliated with Coinbase, recently launched B3IQ, allowing users to rent-to-own servers equipped with NVIDIA GPUs (with a 30% down payment followed by installment payments). During the payback period, users and the platform share computing power revenue in a 3:7 ratio. After full payment, if users choose to continue renting out their computing power, the revenue-sharing structure reverses. While users rent out their computing power, the platform also earns concurrent revenue.

More important than rental prices is the repair of expectations regarding the discounted future value.Based on calculations using the 36-month forward rental curve, Silicon Data indicates that the estimated residual value of a single H100 GPU has rebounded from approximately $14,000 last October to around $20,000 currently. Although this is merely a theoretical valuation rather than an actual transaction price in the secondary market, it indirectly boosts financial institutions' confidence in treating GPUs as collateral for financing:When a piece of equipment can consistently generate stable cash flows, it establishes a foundation for capitalization.

The recovery of the rental curve aligns precisely with the "computing power assetization" business model proposed by NVIDIA CEO Jensen Huang.

Under the traditional capital expenditure model, data center construction heavily relies on cloud providers' cash flows or unsecured corporate bonds, with expansion speed constrained by their balance sheets. Jensen Huang's proposition is:Since GPUs can continuously generate inference rental income throughout their lifecycle, data centers possess characteristics similar to productive assets such as commercial aircraft and large cargo ships. This allows them to leverage future cash flows as collateral to secure leveraged financing.

On August 10, this concept moved into the implementation phase. NVIDIA announced that it had signed memorandums of understanding with six leading financial institutions—Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR—to jointly establish a computing power financing platform aimed at raising over $500 billion in third-party capital for global AI infrastructure development.

In the design of this financing structure, some loans will be directly secured by the purchased GPU equipment and physical data center facilities. To alleviate financial institutions' concerns about rapid hardware depreciation, NVIDIA has committed to providing residual value support or credit enhancement guarantees of up to 25% of the equipment's residual value in specific cases.

Neo-Cloud providers can then repay their debts using future revenue from computing power leases, creating a closed financing loop: "collateralized financing → GPU procurement → computing power leasing → debt repayment."

In fact, there are already pioneers in this model. In August 2023, CoreWeave pledged its H100 cluster to Blackstone and Magnetar Capital, securing $2.3 billion in debt financing.

On August 12, CoreWeave's financial data revealed that its Q2 results comprehensively exceeded expectations, validating the surge in market demand for computing power. Since the earnings release, its stock price has cumulatively risen by approximately 18%. Since its IPO last March, CoreWeave's stock price has actually increased by more than 1.65 times.

Returning to the $500 billion financing platform led by NVIDIA, its significance lies in transforming case-by-case exploration into an industry-wide standardized financing tool:It will significantly reduce financing costs for Neo-Cloud providers, attract more capital into the AI infrastructure sector, and in turn feed back into GPU order demand.

The launch of this financing platform could become a watershed moment for AI infrastructure, marking a shift from a "heavy-asset business" model to "financialized expansion."NVIDIA sells more chips, financial institutions earn returns on AI infrastructure investments, and Neo-Cloud providers gain ammunition for business expansion; together, these three parties keep the flywheel of computing power assetization spinning.

Beneath the $500 billion financing narrative, disagreements have never ceased. The focal point of these disputes strikes at the foundation of computing power assetization: How long is the true economic life of GPUs? Are residual value estimates really reliable?

Bearish view: Hardware depreciation far exceeds expectations

The bearish argument stems from actual transaction data in the secondary market and adjustments in corporate accounting records.

Secondary trading data from Silicon Data shows that used H100s with around three years of usage are listed at only 20%-30% of the peak price of new units. Moreover, liquidity in the secondary market is limited, and actual transaction prices often involve further discounts. From a liquidation perspective, the rate of hardware depreciation for GPUs remains high.

Corporate actions are also reinforcing these concerns. Early last year, Amazon shortened the depreciation period for its servers and network equipment from six years to five. This single change added $1.4 billion in depreciation expenses to its financial statements, reflecting the reality of accelerated AI hardware iteration cycles. Noted short-seller Michael Burry has explicitly stated that major cloud providers' financial reports significantly underestimate the depreciation rate of AI chips, hiding substantial risks of asset devaluation.

Conservatives argue that the 'book residual value,' derived by discounting future rental income, is fundamentally different from transaction prices in the secondary market. Once a generational leap in technology occurs, the collateral value of high-end GPUs could collapse rapidly.

Bull Case: Inference Demand Reshapes Economic Lifespan

Proponents point out a common second-order misjudgment in the market: the widespread confusion between the economic value of GPUs in 'training' versus 'inference' roles.

Admittedly, next-generation architectures like Blackwell will impact older architectures on the training front, pushing legacy chips out of the high-end training market. However, the explosion in inference demand driven by commercial deployment is significantly extending the economic lifespan of GPUs.

The A100 is the most typical example. Six years after its launch, it has long exited the high-end training market, yet it maintains stable leasing rates and robust demand in scenarios such as model inference, low-cost fine-tuning, and domain-specific computing. Its unit economics and practical utility have not been significantly eroded. This implies that older GPU models will not be directly obsolete by new generations; instead, they will cascade down to the inference market, continuing to generate cash flow.

Simply put, the 'true economic lifespan' of GPUs is far longer than the 2–3 years assumed by the market. Sustained growth in inference demand will offset the nominal depreciation caused by technological obsolescence.

The financialization of compute assets is entering a critical phase, serving either as fuel for industrial acceleration or as a source of systemic risk. The rebound in GPU rental rates provides short-term support, while inference demand offers long-term growth potential. However, the pace of technological iteration remains a Damocles' sword hanging over asset valuations. Consequently, the financialization of compute assets could become both a new lever for expanding AI infrastructure and a new channel for risk transmission.

In this joint gamble between Silicon Valley and Wall Street, whoever can more accurately measure the economic lifespan and discounted residual value of compute power will hold the pricing power for the next phase of AI infrastructure.

33K Views