Nvidia, the Guarantor Behind the GPU

After stepping back from trying to compete with major clouds, Nvidia disclosed $3.5 billion in guarantees for other companies’ data-center leases in February 2026. It then reportedly offered to rent young providers’ unused GPUs.

Nvidia is becoming an underwriter and market maker for AI capacity rather than simply its equipment supplier: it can help operators finance land, power and facilities, absorb some risk of spare capacity, and share in the revenue those operators earn. That makes specialized clouds easier to build and gives Nvidia another channel for its systems. It also makes demand for those systems harder to separate from the financial support that helped put them in service.

DGX Cloud showed the limits of selling the service yourself

Nvidia launched DGX Cloud in 2023 as a way for companies to scale AI work without assembling their own systems. A single node started at $36,999 a month. Nvidia was not merely selling the accelerator; it was offering access to a working machine, with the deployment problem already solved for the buyer.

By September 2025, Nvidia had reportedly scaled DGX Cloud back primarily to internal research and development after initially envisioning it as a competitor to major clouds such as AWS. That retreat matters because Nvidia’s subsequent financing activity need not be read as another attempt to own every customer relationship. An operator can lease a building, install Nvidia systems, sell compute and manage customers while Nvidia supports the parts of the project that the operator cannot readily finance alone.

Between 2024 and 2026, Nvidia drew more enterprise-focused coverage: the share of its stories framed around enterprise rose from 15.2% to 24.8%, while the consumer share fell from 18.3% to 10.0%. Those shares describe coverage, not Nvidia’s revenue. The contracts explain why the company increasingly appears in accounts of construction, leasing and credit alongside accounts of chips.

A lease guarantee puts the chipmaker behind the building

An AI infrastructure project needs a site, power and a facility before anyone can install a GPU. Developers must commit to those inputs long before they know exactly how many hours customers will buy from the finished installation. A young cloud operator with little credit history can therefore struggle to obtain the leases that would allow it to purchase and deploy equipment.

In February 2026, Nvidia disclosed $3.5 billion in guarantees for companies leasing data-center land, power and facilities, four times its Q3 level. A guarantee is not the same as an immediate $3.5 billion cash outlay. It is a contingent obligation that can give a landlord or other counterparty a stronger name to rely on if the operator cannot meet its commitments. Nvidia’s balance sheet reaches into the construction conditions behind a hardware sale.

Equity fills a different gap. Applied Digital, a data-center developer expanding into Nvidia-based cloud computing, raised $160 million from Nvidia and other investors. Nebius raised $700 million from Nvidia, Accel and others in December 2024. Investors can fund expansion through shares; lease guarantees address a counterparty’s concern about payment. Treating both as “investment in AI” conceals which obligation each arrangement is meant to carry.

Idle chips give the vendor a second place in the contract

Nvidia reportedly agreed in 2025 to rent 10,000 of its own AI chips from Lambda for $1.3 billion over four years. In July 2026, reporting described a further arrangement: Nvidia would rent unused GPUs from young providers including Firmus and Sharon AI in exchange for a share of their cloud revenue.

Lambda’s reported contract makes Nvidia a buyer of capacity supplied by a company using its chips. The Firmus and Sharon AI arrangement, as reported, goes further toward the operating risk: Nvidia would take capacity that the providers had not placed with other customers. Neither fact establishes that those providers lack independent demand. Each does establish that a chip order and a separately sourced end-customer workload are not necessarily the same transaction.

A lender financing a toll road can distinguish a traffic forecast from a contract that pays when traffic disappoints. The reported GPU rentals create a comparable distinction for a powered rack. Nvidia’s support can make an operator’s obligations easier to carry without itself proving how many outside customers will use the rack. That is the boundary in circular AI financing that deserves inspection, not an assumption that every supported sale is fictitious.

Ohio shows that even a large backstop has a negotiable edge

Nvidia’s proposed support for an Ohio campus illustrates how much can change between a financing plan and a signed commitment. In August 2026, sources said Nvidia had reworked an arrangement so it would initially guarantee only half of a planned $250 billion backstop for an OpenAI data-center project. The revision cuts against any notion that Nvidia simply accepts every risk needed to get capacity built.

A later filing described a different, concrete set of commitments: Nvidia agreed to spend up to $105 billion supporting SB Energy’s Ohio campus, which OpenAI is set to lease, and to invest $1.5 billion in SB Energy. The proposed $250 billion backstop and the filed support agreement should not be added together or treated as identical obligations. Their differences are the point. The developer, prospective tenant and chip supplier can each occupy a separate place in the capital structure, and the amount each bears depends on negotiated terms.

Long-duration leases elsewhere in the buildout show why those terms command attention. Oracle signed about $150 billion in data-center leases in the three months ending November 2025, taking its total data-center and cloud-capacity commitments to $248 billion. A customer’s long-term capacity commitment can support financing; it does not remove the need to establish which party pays if delivery, utilization or asset values fall short of the original plan.

Nebius has outside customers, not an exemption from utilization risk

Nebius reported $575 million in Q2 cloud revenue, up 514% year over year, and $582 million in total revenue. Reflection said it had signed a contract worth more than $1 billion for Nebius computing capacity, including access to Nvidia chips. Nvidia’s investment in Nebius therefore cannot, by itself, explain away the operator’s reported cloud sales or its named outside customer.

Those sales answer one question: customers other than Nvidia are willing to contract for capacity. They do not answer what Nebius earns after serving those workloads, how much of its fleet sits idle, or what its hardware will be worth against outstanding obligations. Lambda, another Nvidia-backed operator, was reportedly seeking up to $4 billion ahead of a planned IPO. Its access to prospective equity capital likewise says something about investors’ appetite, not the profitability of each installed GPU.

Operators also have work Nvidia cannot do for them merely by supplying chips. They must win workloads, place them on available machines, serve requests at a competitive cost and collect enough revenue to maintain their facilities and repay capital. Nvidia can support the channel without making its participants captive resellers—or relieving them of those tasks.

Inference makes placement and resale part of the asset

Training capacity can be sold as access to a large block of machines for a defined job. An inference operator must also manage a stream of requests whose timing and computing needs vary. Nvidia’s Groq 3 LPX inference accelerator entered full production with Nebius as its first customer, putting a new serving product inside the same operator relationship.

Nebius reportedly acquired Inferize for $100 million to $150 million; Inferize’s technology helps optimize GPU utilization and reduce the cost of AI requests. The acquisition addresses a problem that another chip purchase cannot solve on its own. An operator has to turn installed capacity into useful, paid work often enough to support the asset beneath it.

Lenders have begun testing how to value that asset. General Compute received a $400 million loan described as seemingly the first to use inference-specific chips as collateral. A lender taking chips as security has to consider more than their purchase price: it needs a view of what the equipment could earn in use or recover if the borrower fails. Nvidia’s reported promise to take idle capacity has economic weight because it addresses that uncertain interval between installation and a paying workload.

Credit enhancement does not make credit risk disappear

AI infrastructure companies borrowed more than $100 billion in 2025, yet smaller companies faced higher interest rates as debt investors remained wary of unproven businesses. Nvidia’s own commitments to component suppliers reached $279 billion in Q2, up from $50.3 billion in Q3 2025. The chipmaker has obligations upstream of delivery as well as commitments to support projects downstream of it.

Nvidia was reportedly in early-stage talks with insurers about products to protect lenders against neocloud loan defaults. No completed insurance product follows from those talks. They identify another possible holder for a loss that a lease guarantee or capacity rental cannot abolish.

The disclosed evidence does not allocate minimum-utilization payments, termination costs, residual-value losses and defaults across Nvidia, operators, customers, lenders and insurers. Without those contract terms—and without provider-level utilization and margins—nobody can measure how much demand Nvidia’s support creates rather than enables. The distinction comes down to a leased building with power at the door, Nvidia systems in the racks, and a contract naming who pays when those racks are quiet.