Nvidia Holds the Lease

Nvidia is letting other companies’ processors into racks built around its GPUs. It is also agreeing to take cloud capacity that its customers cannot sell. One decision opens the machine; the other puts Nvidia’s balance sheet behind the building that houses it.

Nvidia is becoming a balance-sheet organizer of the AI factory, not simply its dominant chip supplier: NVLink Fusion can keep mixed-silicon racks connected through Nvidia’s interconnect, while investments, leases and capacity guarantees position the company to influence which projects get financed and who bears their idle hours. That creates leverage over AI infrastructure without establishing that Nvidia owns every chip, controls every contract or can profit from every guarantee.

Nvidia now draws more enterprise-focused than consumer-focused coverage. Across the 2024–26 record, enterprise framing rose from 15.2% to 25.1% of Nvidia coverage, while consumer framing fell from 18.3% to 10.3%. Those are measures of coverage, not of Nvidia’s revenue. They mark the distance between a company discussed chiefly for the devices that use its chips and one increasingly discussed through the projects that must house, power and pay for them.

Nvidia first made GPUs easier to deploy

The earlier platform answered a software question. In 2018, Nvidia said it would support Kubernetes orchestration on its GPUs and contribute GPU enhancements to the open-source community. In 2019, Nvidia said it would bring CUDA acceleration to Arm CPUs and share its AI and high-performance-computing software with Arm. Both moves reduced the work a customer had to do before putting Nvidia hardware to use.

That design made sense when adoption was the bottleneck. Developers needed software, operators needed familiar deployment tools, and chip buyers needed confidence that applications would run after the hardware arrived. Nvidia could widen access to those complements while keeping CUDA and its accelerators at the center of the purchase.

A rack-scale data center asks a different question. Its operator must connect CPUs, accelerators and networking into a working system, then keep that system busy enough to justify the building, power connection and financing. Nvidia’s software advantage still matters, but it no longer answers every decision between ordering a chip and earning revenue from it.

NVLink Fusion admits rival silicon without making the rack neutral

In May 2025, Nvidia unveiled NVLink Fusion to couple non-Nvidia CPUs or accelerators with Nvidia GPUs in rack-scale systems. Arm subsequently said Neoverse CPUs would be able to integrate with accelerators through Fusion. SiFive said in January 2026 that it would integrate the technology with its RISC-V processor IP so its silicon could communicate with Nvidia and partner chips.

Nvidia does not have to supply every processor to remain important to the system design. A customer can choose another CPU, or consider a specialized accelerator, while still choosing an interconnect and architecture organized around Nvidia. Opening that connection expands the set of components a buyer can place in the rack; it does not turn NVLink into an independent standard or prove that every participating design will reach deployment.

The physical distinction matters. A data-center operator does not buy “heterogeneous compute” in the abstract. It orders components that must exchange data at useful speeds inside a finite rack, with a cooling system and a power budget already specified. The company that defines those connections can remain difficult to displace even when another company supplies a valuable chip.

Inference is splitting the work that one accelerator used to claim

Training rewarded general-purpose accelerator fleets. Serving a model exposes different constraints at different moments. The prefill phase processes the incoming prompt; decoding generates the response. Nvidia’s Rubin CPX architecture emphasizes compute capacity over memory bandwidth for prefill. Coverage of Nvidia’s Groq agreement has suggested that Groq’s technology could support an ultra-low-latency variant for other inference work, but that remains a possibility, not an announced product.

Nvidia and Groq entered a confirmed $20B, non-exclusive inference-technology licensing agreement, and Groq senior executives moved to Nvidia as part of the arrangement. A license gives Nvidia access to an alternative serving approach without proving that it has bought Groq or secured exclusive control of its technology. The distinction is central: Nvidia can add an option to its own system design even as other companies remain able to pursue specialized inference.

Model builders are doing so. OpenAI and Broadcom unveiled Jalapeño, an LLM-optimized inference chip taken from design to manufacturing tape-out in nine months. Inferact raised $150M around the open-source vLLM inference engine, while DeepInfra reported support for more than 190 open models. Hardware specialization and software portability give buyers ways to question the cost of a default GPU serving stack. The deeper account of inference infrastructure is therefore a story about different jobs inside the same service, not one replacement chip waiting to win them all.

Nvidia is putting capital beside the hardware order

On October 1, Nvidia completed the final $10B tranche of its $30B OpenAI investment pledge. Nvidia also purchased 214.7 million Intel shares at $23.28 each, about $5B. As of late July, Nvidia reported $47.9B held in private companies and $18B committed to equity investments for the remainder of its fiscal year. It sold $25B of high-grade bonds after initially targeting roughly $20B.

Those transactions have different purposes and different limits. OpenAI is a major compute buyer; an equity investment ties Nvidia to the customer’s prospects without itself constituting a hardware order. The Intel purchase is a stake in a chipmaker, not control of its manufacturing or product roadmap. Nvidia reportedly tested Intel’s 18A process but did not proceed with production, so the stake does not establish a new Nvidia manufacturing route.

Nvidia’s February disclosure brought a more direct project obligation into view: $3.5B in guarantees to companies leasing land, power and data-center facilities, four times its Q3 level. A young operator can have demand for GPUs and still struggle to persuade a landlord or lender to finance the site around them. Nvidia’s credit can help close that gap; how much of a failed project it would have to cover depends on the guarantee’s terms.

The unsold hour has become a contracted liability

CoreWeave’s $6.3B order with Nvidia specifies that Nvidia will purchase cloud capacity CoreWeave cannot sell to other customers through April 13, 2032. CoreWeave gets a backstop for utilization; Nvidia takes an obligation linked to whether other customers fill the capacity. That is more precise than saying Nvidia “supports” a cloud provider. It identifies the hour no one bought and names a party that has agreed to buy it.

Nvidia offered a related arrangement to young providers including Firmus and Sharon AI: it would rent back unused GPUs in exchange for a share of their revenue. The operator can finance and offer capacity with some protection against empty machines; Nvidia participates in the upside while accepting a defined route for unused equipment back onto its books. GMI Cloud’s $668M raise shows another division of roles: $223M came through equity financing with Nvidia participation, while a CTBC-led credit component supplied $445M. Nvidia’s equity participation does not mean Nvidia guaranteed that loan.

These are distinct instruments of AI infrastructure finance. Equity absorbs losses after creditors, a lease guarantee supports a facility obligation, and a capacity purchase pays for service that another buyer did not take. None requires Nvidia to own a cloud operator outright. Each can affect what the operator is able to build and what its lenders are willing to fund. Nvidia has reportedly held early-stage talks with insurers about protecting neocloud lenders against defaults, but no completed insurance product or underwriting business follows from those talks.

Other balance sheets—and cheaper serving—limit the position

Broadcom has agreed to lend Anthropic up to $42B through a convertible note that could help finance a $125.2B, five-year TPU lease commitment. That arrangement supplies a counterexample to any claim that Nvidia is the only route from specialized silicon to financed capacity. Broadcom can support demand for a different compute stack with credit of its own. Nvidia and Broadcom have also used residual-value guarantees to support customer purchases, making the future worth of equipment part of today’s financing terms.

Large commitments are not interchangeable with deployed demand. Nvidia’s announced plan to invest up to $100B in OpenAI reportedly stalled before the companies moved toward the smaller, up-to-$30B structure that Nvidia has now funded. Nvidia also reportedly reworked an Ohio data-center arrangement so that it would initially guarantee only half of a planned $250B backstop. Nvidia was willing to support the projects, but it also negotiated the size and timing of the risk it would assume.

OpenAI and Anthropic documents reported inference costs exceeding half of revenue, which explains why serving efficiency matters to both buyers. OpenAI engineers reportedly found a way to more than halve inference cost. A lower cost per response can help a model company, yet it can also alter how much capacity that company needs to reserve and what existing equipment is worth. Without each contract’s utilization, termination and residual-value terms, neither a headline investment nor a headline guarantee reveals who ultimately takes that adjustment.

A planned Texas data center puts the new arrangement in one place. Anthropic signed a $35B cloud deal with Nvidia-backed Lambda; under the arrangement, Hut 8 is to build the facility, Nvidia will supply its chips, and Nvidia will hold the lease. The builder, cloud provider, customer and chip supplier have separate names. Nvidia’s is the one slated for the lease.