Nvidia Signs for the Building

Nvidia plans to upgrade its AI accelerators every year. Asset managers reportedly want Nvidia to stand behind the resale value of the accelerators those upgrades leave behind.

In February, Nvidia disclosed $3.5 billion in guarantees for companies leasing land, power and data-center facilities, four times the prior quarter’s reported level. The guarantees may help builders without credit ratings secure obligations that a chip purchase alone cannot satisfy. A GPU can be ordered before its operator has proved that customers will keep it busy; the building still needs a lease.

Nvidia is becoming an AI-infrastructure credit counterparty, not just its chip supplier. Lease guarantees support construction obligations, reported agreements to rent back unused GPUs address utilization, and proposed residual-value support would put part of the hardware’s obsolescence risk back on Nvidia. Its ability to finance demand, and to govern the exposures created in doing so, now matters alongside GPU performance and CUDA adoption.

Nvidia built the platform before lenders had to value its used chips

Nvidia’s original commercial question was how to make a processor useful across more workloads. Its 2018 Rapids launch extended GPU acceleration into data science and machine learning with partners including IBM, HPE, Oracle and Databricks. A 2019 collaboration brought CUDA acceleration to Arm CPUs. Software support gave customers reasons to keep using Nvidia hardware after a benchmark result had ceased to be news.

Scarcity strengthened that position. Nvidia held an estimated 95% share of machine-learning GPUs when the A100 was described as a critical generative-AI tool in 2023. Its H100 later led all nine MLPerf 4.0 training benchmarks. By 2025, Nvidia was selling GB200 Blackwell servers containing 72 chips per unit and claiming tenfold performance over H200 systems for certain mixture-of-experts models. Buyers were choosing an integrated system, not merely a replaceable processor.

That history matters to a lender valuing a used rack. A lender cannot rely solely on what the chips accomplished when new; it must judge whether another operator will have compatible workloads and software when the first borrower no longer wants them. Nvidia’s installed base and CUDA ecosystem give that judgment a basis, while annual upgrades threaten the price an older generation might fetch. Nvidia’s NVLink Fusion, which permits non-Nvidia CPUs or accelerators in rack-scale setups with its GPUs, broadens the possible configurations without making every used configuration equally easy to resell.

Nvidia drew enterprise-framed coverage in 15.2% of stories about it in 2024 and 23.8% in 2026. That is a measure of coverage, not customer revenue. The purchase under discussion increasingly includes the rack, its operator and the facility around it.

A residual-value promise answers a different question from a lease guarantee

Nvidia’s disclosed lease guarantees concern obligations attached to land, power and facilities. The company has also reportedly offered young cloud providers a different form of support: renting back unused GPUs for a share of their revenue. That arrangement addresses an empty-rack problem rather than a missed facility payment. The distinction matters because an operator can pay for its building yet fail to earn enough from the machines inside it.

A residual-value guarantee reaches further into the hardware’s life. If a lender expects to recover part of its loan by selling GPUs after a lease ends, a supplier’s support can reduce the lender’s exposure to a disappointing resale price. The Financial Times reported that Nvidia and Broadcom are among the suppliers using residual-value guarantees to support customers’ off-balance-sheet AI purchases. Such support changes the price of credit; it does not create a buyer for every old accelerator.

Asset managers reportedly pressed Nvidia to increase a 25% residual-value support backstop for a planned $500 billion AI-financing vehicle. The vehicle and backstop remain sources-based plans, and the reporting said no deals had emerged from the initiative. The negotiation establishes what prospective financiers wanted Nvidia to bear, not what Nvidia has ultimately guaranteed.

The hardware makes that negotiation unusually sensitive. Nvidia’s new systems combine chips with interconnect and software choices that an eventual buyer must be willing to use. A resale estimate is therefore also an estimate of future workload demand and the continued usefulness of Nvidia’s platform. The supplier promising a residual value has considerable influence over the next product generation that will compete with the old one.

Supplier commitments and customer credit can fail together

Vendor support can help a customer raise construction money before it has an investment-grade credit history. It also makes a sale less independent of the vendor’s financial capacity. Reporting put Nvidia’s commitments to component suppliers at $279 billion in Q2, up from $119 billion in Q1. Those purchase commitments are not claims under customer guarantees and should not be added to them as though they were the same liability. They do put Nvidia on another side of the same buildout: components must be secured before every customer has proved its long-run utilization.

Nscale’s filing shows why the distinction between an order and a funded deployment matters. After ordering Nvidia chips for a $45 billion Anthropic deal, Nscale said it needed more funding for its 2027 commitments, casting doubt on a planned $35 billion IPO. The filing does not establish that Nvidia guaranteed Nscale’s obligations. It shows that chip demand can be real while the customer’s financing remains unsettled.

AI-infrastructure companies borrowed more than $100 billion in 2025, and smaller borrowers faced higher rates amid doubts about unproven businesses. By 2026, CoreWeave bond yields had topped 15%, while some Oracle debt traded near 80 to 90 cents on the dollar. A vendor backstop can change who absorbs a loss; it cannot make GPU rental prices, customer credit and resale values move independently. Nvidia’s broader capital-allocation role makes the boundaries between those exposures worth examining contract by contract.

Google gives financiers another collateral story

Nvidia cannot assume that every future buyer of used AI capacity will want the same hardware. Google assembled a reported roughly $200 billion Anthropic financing program, with more than $150 billion tied to TPUs and participation from Broadcom, Blackstone, Apollo and others. Broadcom was separately reported to be assembling $60 billion in AI-chip financing for Anthropic and other customers. Neither program proves that TPU-backed equipment will hold its resale value better than Nvidia systems. Both give financiers another package of chips, customers and capital to evaluate.

Other alternatives narrow the assumptions Nvidia can make about a secondary market. Cerebras reported 70% year-over-year growth in Q3 core revenue to $230.9 million and announced a Microsoft deal. Huawei has positioned its Atlas 950 and 960 SuperPoD systems as rivals to Nvidia’s accelerators. These suppliers need not displace Nvidia across the market to affect an appraisal: a lender values a particular fleet against the workloads and competing capacity available when that fleet must be sold.

Nvidia’s board now owns the collateral file

Jensen Huang resigned as Nvidia’s CEO, president and board member on November 12, and CFO Colette Kress became interim CEO while the board considers candidates. Nvidia appointed former Synopsys CEO Harvey C. Jones as its first board chair. Their immediate governance task extends beyond choosing who directs chip development. The board needs to know which customers receive guarantees, which contracts could require Nvidia to take back capacity, and how the company tests resale assumptions before accepting them.

Export compliance belongs in that file. A Reuters review identified more than $3.2 billion in Chinese compute projects built or planned around Nvidia GPUs restricted by the US. That review does not establish that Nvidia knowingly supplied a prohibited project. It does show why a financier assessing collateral needs a credible chain of custody: where the chips are, who can use them and whether a future sale is legally available all bear on what a guarantee might cost. US officials have also said multiple Justice Department investigations could sweep up Nvidia, including an antitrust-focused inquiry and investigations involving Megaspeed and Supermicro.

Kress and the board inherit promises made across sales, supply, credit and compliance decisions that cannot be assessed one desk at a time. Nvidia can replace a chip generation; it cannot upgrade away its signature on a data-center lease.