NVIDIA’s Hundred-Billion-Dollar Bet on “AI Factories”: From Selling Chips to Building a Computing Empire

On August 17, 2026, NVIDIA dropped a bombshell—announcing a $1.5 billion investment in AI infrastructure developer SB Energy and providing up to $105 billion in credit guarantees for the PORTS-Pike technology park in Ohio. This hyperscale data center will be dedicated to deploying NVIDIA’s AI computing infrastructure, with OpenAI as the sole tenant.

This is not just a transaction with an eye-popping sum and a complex structure; it marks a shift in global AI competition from mere hardware procurement to a new era where “infrastructure as an asset” takes center stage. As computing demand begins to strain the limits of traditional financing models, NVIDIA has chosen to step in directly, using its own balance sheet to pave the way for next-generation AI infrastructure.

The Chess Game Behind the Billion-Dollar Guarantees

According to the agreement, the PORTS-Pike campus is located in Pike County, Ohio, adjacent to the decommissioned Portsmouth Gaseous Diffusion Plant site. The land comprises both private parcels and federally controlled areas, benefiting from industrial zoning policies and access to substantial power capacity. The initial deployment capacity is 4.25 IT-GW (gigawatts), and NVIDIA holds an option to expand by an additional 3.75 IT-GW, bringing the total planned capacity to 8 IT-GW.

In terms of hardware scale, NVIDIA disclosed that each generation of AI factory systems deployed at the campus will contain approximately 1.5 million GPUs. Based on current cost estimates for AI computing clusters, these 1.5 million GPUs correspond to potential revenue opportunities of roughly $150 billion to $200 billion. It is worth noting that such massive computing power will not be deployed all at once; rather, it will be rolled out progressively in line with NVIDIA’s GPU generational updates, aligned with OpenAI’s model training and inference demands.

Breaking down the business model, this is a carefully orchestrated, risk-sharing tripartite arrangement: OpenAI bears the long-term leasing obligation and actually uses the compute power, serving as the ultimate anchor on the demand side; SB Energy handles land development, civil construction, and daily operations, playing the role of a traditional infrastructure operator; NVIDIA provides direct investment and up to hundreds of billions of dollars in credit guarantees, offering financial backing and core hardware supply. According to subsequent NVIDIA disclosures, the company may also provide up to 25% residual value support for individual investment projects—a mechanism designed to allay financiers’ concerns about rapid GPU depreciation due to technological obsolescence.

In fact, NVIDIA had earlier adjusted its total guarantee amount from the market-rumored $250 billion down to no more than $120 billion. NVIDIA explicitly stated that the project does not involve revolving financing, and the ultimate lease payment responsibility falls on OpenAI. This structural design keeps NVIDIA’s credit risk exposure relatively manageable while still leveraging billions in external capital for AI infrastructure.

OpenAI’s commitment is equally staggering. Under the compute leasing arrangement, OpenAI plans to deploy approximately 12 gigawatts of NVIDIA computing power by 2030. If the expansion options at the PORTS-Pike campus are fully exercised, the deployment scale could rise to about 16 gigawatts. Based on this long-term roadmap, NVIDIA estimates that the related computing opportunities could correspond to roughly $600 billion in NVIDIA computing revenue by 2030—a figure several times NVIDIA’s current annual revenue, reflecting extreme optimism about long-term AI compute demand from both parties.

From “Selling Chips” to “Building Factories”: The Deep Strategic Shift

The significance of this partnership goes far beyond setting a new industry record for transaction size; it reveals a profound repositioning of NVIDIA’s strategic priorities.

Over the past few years, the core resources contested in the AI industry have been GPU compute chips, high-bandwidth memory (HBM), advanced packaging capacity, and high-speed interconnect networks. But as large-model parameters continue to scale toward trillion-level magnitudes and generative AI inference frequencies grow exponentially, the traditional model of “buying chips and renting server space” is hitting new physical limits. Today, land ownership, substation capacity, cooling water sources, and multi-year data center construction cycles are becoming the new bottlenecks constraining computing expansion.

NVIDIA defines AI factories as the “defining infrastructure of the intelligent era,” and ranks land, power, and plant facilities as strategic resources on par with chip design, packaging, memory bandwidth, and networking. Its supply-chain logic mirrors that of traditional manufacturing giants: when customer demand is highly visible, lock in critical production inputs and factory capacity in advance. This time, however, NVIDIA’s “inputs” are not silicon wafers or chemical reagents, but square kilometers of industrial land and megawatts of grid capacity.

Analysts generally point out that this signals the chip giant’s evolution from a pure AI hardware vendor to a core participant and master architect of “AI factory” infrastructure. NVIDIA is building a complete AI factory delivery system spanning chips, networking, software stacks, and physical infrastructure. For customers, renting NVIDIA’s computing power in the future might be as convenient as leasing standardized factory space today—turnkey and scalable on demand.

Behind this strategic pivot is the reality that cutting-edge AI companies’ compute demands are rapidly exceeding the limits of their own balance sheets. NVIDIA has internally noted that the computing growth rates of some frontier AI labs have already outstripped their independent financing and construction capabilities. The PORTS-Pike agreement offers a novel solution: NVIDIA leverages its strong credit rating to guarantee infrastructure financing, while OpenAI locks in compute capacity for the next decade through long-term operating leases, avoiding a one-time capital expenditure of tens of billions that would strain its cash flow.

Market Jitters and Jensen Huang’s Reassurance

Such a massive and structurally innovative capital move naturally drew intense attention on Wall Street and in tech circles, along with considerable concern.

Just a week earlier, NVIDIA CEO Jensen Huang had joined forces with six top Wall Street asset managers to announce plans to turn AI computing into a standalone, tradable asset class. Market sentiment was initially buoyant, but skepticism quickly emerged over the debt risks implicit in the “revolving financing” structure and the potential strain on NVIDIA’s own balance sheet. In response, Huang personally stepped in to soothe investors, emphasizing that NVIDIA would prudently evaluate every infrastructure investment project, strictly control guarantee exposure, and noted that the residual value support mechanism would effectively mitigate counterparty default risk. After these remarks, market anxiety gradually eased.

Looking at capital market reactions on August 18, AI computing stocks showed clear divergence. The communications sector, which had experienced a rapid recovery, entered a consolidation phase; the communications ETF pared gains but still closed down about 1.34% in the afternoon, indicating growing short-term divergence among funds regarding high-valued computing stocks. However, North American cloud service providers’ shares held relatively firm, with the market generally believing that NVIDIA’s move would accelerate the standardization and financialization of AI compute, benefiting the entire industry chain’s valuation over the long term.

Chinese Computing Power: Resonating in Sync, But on a Different Path

In this global AI infrastructure race led by NVIDIA, China’s AI computing industry has not remained on the sidelines—it is moving in resonance, albeit with distinctive characteristics.

Around the same time, domestic computing power is accelerating its shift from “single-card performance competition” to “system-level cluster scaling.” Huawei’s Ascend Atlas 950 SuperPoD made its first live appearance at WAIC 2026, with the system capable of horizontally scaling to 8,192 Ascend cards in high-speed interconnection, targeting training and inference for trillion-parameter models, with mass production scheduled for Q4 2026. Alibaba Cloud introduced the Pangu AL128 super-node server, densely integrating 128 self-developed Zhenwu M890 chips in a single rack, deeply optimized for elastic computing demands in public cloud scenarios. Sugon (Sugon) also announced the completion of China’s first domestic 100,000-card AI supercluster, marking a key step toward self-reliance.

Policy support is also accelerating. On August 18, six departments in the Inner Mongolia Autonomous Region jointly issued measures to support enterprises in building ultra-large-scale computing centers using domestically produced hardware, with guarantees in land use, green power quotas, and energy-consumption targets. Shortly after, Sichuan province unveiled 19 policy initiatives to advance its western computing hub, covering electricity subsidies, network latency optimization, and computing resource grid integration.

Demand-side data is even more telling. According to the latest OpenRouter statistics, China’s weekly AI large-model invocation volume reached an astonishing 34.25 trillion tokens, up 21.76% month-on-month. This marks the 15th consecutive week that China has surpassed the United States, ranking first globally—showing that the explosive growth of domestic AI applications is placing unprecedented and urgent demands on underlying Chinese computing infrastructure.

The Era of Compute as an Asset Has Fully Arrived

NVIDIA’s hundred-billion-dollar gamble reveals an irreversible trend: computing power is evolving from a continuous “operational consumable” into a long-term, value-preserving “core asset” with financing attributes. When a chip design company is willing to commit hundreds of billions of dollars in credit lines to secure barren industrial land and supporting power facilities, it is betting not merely on the sales of the next-generation Blackwell or Rubin architecture GPUs, but on humanity’s seemingly insatiable, near-infinite demand for AI computing over the next decade.

Jensen Huang’s grand vision, in partnership with Wall Street titans, of turning AI compute into an independent asset class is moving from slide-deck blueprints to a tangible reality—with physical carriers, cash-flow backing, and financial leverage. For the entire global AI industry chain, this paradigm shift from “selling chips” to “building factories,” along with the accompanying wave of compute financialization, may have only just begun to crest. In the coming years, whoever can first close the complete loop—from chip design, hardware manufacturing, and networking, to infrastructure financing and power security—will be best positioned to claim the most advantageous ground in the battle for supremacy in the intelligent age.