NVIDIA's Compute Leasing Business: From Selling Chips to a Strategic Leap as the "AI Central Bank"

I. Business Overview: From DGX Cloud to the "AI Compute Partnership"

NVIDIA's compute leasing business has undergone a clear strategic evolution.

Phase One: DGX Cloud – An Exploration of First-Party Cloud Services (2023–2025)

In 2023, NVIDIA launched DGX Cloud, positioning it as a direct‑connect AI cloud service for enterprises, with monthly rental fees for each H100 instance reaching as high as $36,999. At that time, severe GPU shortages allowed the high‑pricing strategy to hold. However, in the second half of 2024, as GPU supply eased, major cloud providers such as AWS significantly cut prices, with H100 and A100 instance prices dropping by as much as 45%, eroding DGX Cloud's price advantage. By mid‑2025, NVIDIA's Q2 FY2026 earnings report no longer attributed billions of dollars in cloud spending commitments to DGX Cloud, and the service quietly pivoted to internal infrastructure and R&D use. Replacing it was Lepton, launched in 2025—a GPU leasing and scheduling marketplace that connects developers to GPU compute resources from global cloud providers, rather than providing compute directly.

Phase Two: The "AI Compute Partnership" – From Selling Products to Building an Ecosystem (July 2026)

On July 1, 2026, NVIDIA announced a new AI infrastructure partnership model. Its core mechanisms include:

• Revenue sharing: While receiving hardware sales revenue, NVIDIA shares in the cloud service revenue with AI cloud providers, creating a recurring income stream tied to compute usage.

• Credit support and buyback backstop: If cloud providers fail to find enough customers to rent compute capacity, NVIDIA will lease back unused GPU capacity at an agreed price. In exchange, NVIDIA takes a percentage of the cloud provider's revenue, with the share gradually decreasing over the contract term.

• No direct GPU leasing: NVIDIA does not directly rent compute to startups. Instead, it provides GPUs, software platforms, and credit support, while partners such as Sharon AI and Firmus convert these resources into cloud‑based compute services.

Among the first partners, Sharon AI plans to deploy up to 40,000 GB300 GPUs; Firmus is building a DSX AI factory campus in Batam, Indonesia, with plans to scale to 360 MW of power capacity and deploy up to 170,000 NVIDIA GPUs. Together, the two projects could deploy up to 210,000 GPUs.

This model is internally referred to as the "AI Compute Partnership." According to disclosures by SemiAnalysis, NVIDIA's broader initiative is the "Backstop" plan—leveraging its own AA/Aa2 investment‑grade credit rating to provide minimum revenue guarantees for AI compute lessors for up to six years. If third‑party demand falls short, NVIDIA purchases the compute at preset prices; if capacity is leased at higher rates, NVIDIA can share in the excess returns.

II. Strategic Intent Analysis

NVIDIA's moves are far from simple business expansion; they represent a multi‑layered strategic play.

First, reducing dependency on top cloud customers. Currently, a handful of large cloud providers—Amazon, Microsoft, Google, and others—buy the bulk of NVIDIA's chip production. However, several of these are developing their own competing AI chips. To reduce reliance on these giants, NVIDIA has spent years nurturing emerging GPU cloud providers like CoreWeave, and the "AI Compute Partnership" is a continuation and deepening of that strategy. By fostering new compute suppliers, NVIDIA is building a more diversified customer base.

Second, evolving from a chip supplier to the "central bank" of the AI compute ecosystem. NVIDIA is turning its strong balance sheet into a market lever. By providing financial backing to emerging cloud providers in exchange for revenue sharing, NVIDIA is no longer just a "pick‑and‑shovel" seller of chips; it is deeply embedding itself in the profit distribution of the downstream compute market. The "Backstop" plan goes further—using its own credit rating to catalyze bank lending, serving as a financial engine for the expansion of the entire AI infrastructure. Some analysts note that NVIDIA is acting like the "central bank" of the AI era, using its AA‑rated credit to leverage trillions of dollars in AI‑related debt.

Third, seizing the compute gateway in the era of "AI factories." NVIDIA CEO Jensen Huang stated at the 2026 annual shareholder meeting that data centers are transforming from places that store and retrieve information into "AI factories" that produce digital intelligence. AI has evolved from generative AI to agentic AI, and the logical loop—"AI can do useful work → tokens have value → tokens generate profits → compute demand continues to accelerate"—is already in place. Under this vision, whoever controls the distribution of compute holds the key to the entire AI industry chain. Through the "AI Compute Partnership," NVIDIA embeds itself across the full value chain from chips to cloud services, creating multi‑level alignment of interests.

Fourth, creating recurring revenue streams. Under the traditional model, NVIDIA's revenue was heavily dependent on one‑time chip sales, leading to pronounced cyclicality. Through revenue sharing, NVIDIA gains a recurring income stream tied to compute usage. This helps smooth revenue volatility and enhances valuation stability.

III. Impact on the Compute Market

(1) Compute lease prices continue to rise

The launch of NVIDIA's new model coincides with extreme supply‑demand tightness in the compute market. According to SemiAnalysis, one‑year lease contract prices for H100 GPUs surged from $1.70 per hour in October 2025 to $2.35 per hour in March 2026, an increase of nearly 40%, with on‑demand rental capacity for all GPU types completely sold out. The average on‑demand rental price for B300 GPUs jumped over 50% from November 2025, and some cloud providers plan to raise B200 renewal prices by approximately 94%. NVIDIA's "backstop" mechanism effectively sets a floor under lease prices, further reinforcing market expectations of price hikes.

(2) Accelerating the transformation of compute supply models

Compute supply is shifting from the traditional heavy‑capex model to a hybrid structure of "capex plus off‑balance‑sheet financing." NVIDIA's credit backing enables emerging cloud providers, which previously struggled to obtain bank loans, to leverage larger amounts of capital. One data center executive commented that NVIDIA's backstop commitment "kills two birds with one stone"—solving both GPU financing and data center financing. This is likely to accelerate the global expansion of AI compute infrastructure in the near term.

(3) Reshaping industry competitive dynamics

NVIDIA's strategy of nurturing emerging cloud providers is challenging the dominance of top providers like AWS, Azure, and Google Cloud. At the same time, Meta announced in July 2026 that it would lease out its idle AI compute capacity. The compute leasing market is now seeing participants ranging from specialized cloud providers to chipmakers and tech giants, making the competitive landscape more complex.

However, the flip side is risk: NVIDIA has deeply tied its own creditworthiness to the health of the entire AI compute industry chain. If AI application demand falls short of expectations, a large number of "backstop" commitments could translate into actual losses. SemiAnalysis predicts that global AI‑related debt could exceed $7 trillion by 2029. In moving from "selling shovels" to "financing shovels," NVIDIA gains greater industry influence while concentrating systemic risk heavily on its own balance sheet.

Conclusion

NVIDIA's strategic evolution in compute leasing is essentially a three‑step leap from a "product company" to a "platform company" and then to an "ecosystem financial company." The contraction of DGX Cloud was not a failure but a proactive strategic pivot—rather than competing head‑on with cloud giants in a red ocean, NVIDIA is leveraging its credit and balance‑sheet strength to become the "central bank" of the AI compute ecosystem. The success of this strategy will depend on whether AI inference demand continues to materialize—after all, even the most powerful financial leverage requires real application demand to sustain it.