Jensen Huang’s Case for Treating H100 GPUs as Revenue-Generating Infrastructure

Nvidia CEO Jensen Huang is making a financial argument as much as a technology argument: a GPU should not be judged solely as equipment that loses value with age. If it can be deployed across customers, remain heavily utilized and earn recurring rental income, it can function more like productive infrastructure.

His evidence is the continued market value of the H100, Nvidia’s prior-generation AI GPU. The data supports real demand for older capacity. It does not, by itself, settle the harder question of how long an H100 can sustain attractive cash flows once newer systems become widely available.

The argument: compute can be rented repeatedly

Huang described Nvidia compute as “fungible, durable and highly rentable” and as a “productive, revenue-generating asset.” 16 In practical terms, the claim rests on three ideas:

• Fungibility: capacity can be reassigned among customers and workloads rather than being tied to a single buyer or application.
• Durability: a GPU’s commercial usefulness can persist after a newer architecture launches.
• Rentability: operators can turn the hardware into recurring revenue through cloud and contracted compute services.

That is a different lens from a conventional server-refresh model. An asset may depreciate on an accounting schedule while continuing to generate meaningful operating income. For cloud operators and AI infrastructure providers, the relevant economic test is whether rental revenue, utilization and margins justify the capital invested over the asset’s useful life.

What the H100 pricing data shows

The immediate catalyst for Huang’s comments was an Ornn Exchange price index indicating that a roughly three-year-old H100 rented for $3.28 per GPU-hour after rising 22% over the month. 16 That is evidence that buyers still place value on Hopper-generation capacity even as Nvidia sells newer products.

Longer-duration pricing pointed in the same direction earlier in 2026. SemiAnalysis reported that one-year H100 rental contract pricing rose from $1.70 per GPU-hour in October 2025 to $2.35 in March 2026—an increase of nearly 40%. 61

Those figures are useful demand signals, but they should not be treated as a universal H100 price. Published on-demand quotes in 2026 ranged roughly from $2 to $12 per GPU-hour; interruptible capacity can be cheaper, while premium hyperscaler configurations can be substantially more expensive. 49 The apparent dispersion reflects meaningful differences in contract duration, availability guarantees, geography, networking, support and whether a customer is renting a single GPU, a dedicated node or a larger cluster.

Why rental prices matter to infrastructure economics

A GPU fleet becomes more financeable when an operator can forecast its cash generation with confidence. Higher prices and strong utilization can let an operator recover its capital cost across more profitable operating years. That is the economic logic behind treating compute as infrastructure rather than as a one-time hardware purchase.

But rental pricing is only one input. A durable infrastructure thesis needs evidence that fleets can maintain:

1. High utilization across economic cycles and customer segments;
2. Contract renewals rather than merely short-lived scarcity demand;
3. Healthy margins after power, networking, operations and depreciation; and
4. Residual value once newer GPU generations are broadly deployed.

An H100 can remain technically capable yet still suffer economic obsolescence if customers receive meaningfully better performance per dollar or per token from a newer accelerator.

Nvidia is building a financing case around the thesis

Nvidia’s strategy goes beyond messaging about rental rates. In August, the company announced partnerships with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to establish compute-financing platforms intended to mobilize more than $500 billion in third-party capital for AI infrastructure. 33

The wording matters: the initiative is designed to mobilize capital over time, not to show that $500 billion has already been invested or guaranteed. Still, it shows Nvidia’s ambition to frame compute, data centers and related systems as assets that can support long-duration, usage-linked financing.

Nvidia has also agreed to acquire Hugging Face for $12.93 billion, extending its reach into an open-model and developer platform. 31 The deal does not prove that any individual GPU generation will retain value, but it may strengthen Nvidia’s presence in the ecosystem that creates demand for AI compute.

The risks Huang’s argument must overcome

The strongest rebuttal is not that H100s are useless—they plainly retain demand—but that today’s rental market may not predict tomorrow’s returns.

Several developments could weaken the infrastructure case:

• Performance displacement: New architectures could materially improve cost per token or training performance, reducing what customers will pay for H100 capacity.
• Supply normalization: More available GPUs and data-center capacity could pressure both spot and contracted rental rates.
• Alternative accelerators: Proprietary chips and competing AI accelerators could reduce demand for Nvidia capacity in particular workloads.
• Temporary bottlenecks: A price increase can reflect short-term scarcity, rather than enduring scarcity or a lasting improvement in asset value.
• Operating-cost pressure: Power, networking and data-center costs can reduce returns even when headline GPU rental prices remain high.

This is why a one-month 22% increase is best read as a notable demand datapoint, not a complete answer to the depreciation debate.

The bottom line

Huang’s claim is plausible in a narrow but important sense: older AI GPUs can continue to earn substantial rental revenue, and the H100’s recent pricing shows that customers still value that capacity. 1661 Nvidia’s financing partnerships further demonstrate an effort to turn that operating reality into an investable infrastructure model. 33

The thesis will be tested by what happens after newer systems scale: whether H100 fleets preserve utilization, renew contracts and generate sufficient net cash flow after operating costs. Durable technical usefulness is necessary. Durable economic returns are the proof investors and infrastructure lenders will ultimately require.