Nvidia's $3.5 Billion Bet Behind the Rack

In February 2026, Nvidia disclosed $3.5 billion in guarantees for customers leasing land, power and data-center facilities. It sells the chips inside those facilities; now it backs some of the leases that help get them built.

Cheaper inference broadens who can use AI, but it does not distribute the ability to build or secure frontier capacity. Nvidia, a narrow group of manufacturers and memory suppliers, financiers and governments increasingly determine access through physical capacity and permission to buy—not through the advertised price of a token.

Cheap inference increased the load on a system built for scarce hardware

In 2016, Nvidia’s Pascal launch promised faster, more efficient computing. The company could answer a familiar hardware question: how much more work could a customer get from a new generation of chips? Server partnerships and CUDA-compatible software helped customers put those chips to work. By 2023, an A100 had become a critical tool for generative AI, with Nvidia estimated to hold 95% of the machine-learning GPU market. Product performance still mattered, but the installed software and operating practices around the product mattered too.

Lower inference costs make more uses of AI economical, and those uses can increase total demand for compute. A developer buying a cheaper API, however, does not acquire the accelerators, electricity and long-term capacity behind it.

OpenAI said in April 2026 that it had signed contracts for 10GW of U.S. AI compute capacity, including more than 3GW added in the preceding 90 days. Contracts are not completed data centers. They do show a buyer securing future supply on a scale that token pricing cannot describe. AI infrastructure companies borrowed more than $100 billion in 2025, according to Refinitiv figures reported by the New York Times; smaller firms faced higher rates as lenders weighed unproven businesses. Bain estimated that AI companies would need $2 trillion in combined annual revenue by 2030 to fund projected compute demand and could fall $800 billion short. Bain’s estimate is a forecast, not an operating result, but it locates the risk in the buildout rather than in today’s API bill.

Nvidia drew enterprise framing in 25.0% of its 2026 coverage, up from 15.2% in 2024, while consumer framing fell from 18.3% to 10.5%. Those are measures of coverage, not Nvidia’s revenue mix.

Nvidia began underwriting the customers it once supplied

Nvidia’s $3.5 billion in guarantees was four times the level reported for its third quarter. A guarantee is not the same as a cash outlay: its value to a customer lies partly in helping a company without a credit rating obtain financing. Its significance to Nvidia lies in the obligation it assumes if the customer cannot meet the lease.

Nvidia has also promised to rent unused GPUs from young cloud providers in exchange for a share of their revenue. One arrangement addresses the landlord or lender before construction; the other addresses utilization after the racks are installed. Nvidia can sell hardware into a project, support the project’s ability to obtain capital and absorb some risk that its capacity goes unused. The adjacent GPU cloud capacity contracts market makes that last risk explicit.

Nvidia is not alone in tying chip demand to capital formation. Broadcom arranged a $63 billion financing deal through a special-purpose vehicle for Anthropic chip purchases. Broadcom’s participation makes supplier-supported financing harder to dismiss as an eccentric Nvidia tactic. Lenders still have to judge whether the eventual computing demand will support the debt; a chip order and a financing commitment can reinforce one another without proving that either will earn its projected return.

Packaging and power made the cloud a place, not a price

A working AI system needs more than an accelerator design. TSMC and its partners in Taiwan occupy a central position in advanced chip packaging, which joins components into usable high-performance devices. Intel has said the United States holds only 3% of global advanced-packaging capacity. Adding a rival chip design does not immediately add a packaging line able to manufacture it at scale.

Memory can tighten the same constraint. Nvidia reportedly asked Samsung Electronics, SK Hynix and Micron to use TSMC’s 3nm process for custom NVHBM base dies; Samsung reportedly agreed, in what would be its first use of TSMC for HBM. Even competing memory suppliers can converge on the same manufacturing partner when a chip specification requires it. Optical-component shortages have reached lasers, substrates, fibers and connectors, while U.S. AI labs have deployed on-site gas generators where grid capacity struggles to keep pace.

Those dependencies make switching suppliers a construction and operations decision, not merely a benchmark comparison. A data center commits to a site, an electrical connection, an accelerator fleet and software that runs on that fleet. The AI cloud capacity market therefore has locations and lead times even when customers encounter it as a service. A lower rental price cannot move a packaging plant or energize a campus.

Washington made the chip sale conditional on destination

The Biden administration’s 2025 AI-chip export framework gave 18 key allies full access while requiring licenses for most other countries, including China. U.S. approval of Nvidia H200 exports to China later came with third-party testing of the chips’ AI capabilities before shipment. In neither case could a buyer settle access by agreeing to the price alone: the U.S. Commerce Department determined the terms under which hardware could cross a border.

Suppliers have acquired a role in monitoring what happens after shipment. Nvidia reportedly required Southeast Asian customers seeking Nvidia Cloud Partner status to join a program tracking the locations of their AI chips. That reported condition applies to applicants for a particular status, not necessarily every customer. A provider must account for where its machines are to participate in the supplier’s program. An investigation also found that some Chinese firms approved by Commerce to handle H200 chips had empty offices and links to sanctioned companies, exposing how much any permission system depends on verifying the recipient.

China is building a different route rather than waiting only for U.S. approvals. Huawei supplies compute hardware while DeepSeek contributes model workloads and kernel optimization; DeepSeek plans to use Huawei’s Ascend 950DT for pre-training. The work is substantial, but domestic substitution does not erase component constraints. Huawei has reportedly told clients that a large share of its next flagship AI chip will rely on HBM from Samsung and SK Hynix. Export rules can spur an alternative ecosystem while that ecosystem remains dependent on scarce inputs of its own.

Rival chips can loosen Nvidia’s grip without shrinking the buildout

Technical competition is real. AMD has launched competing accelerators, although benchmarking found software bugs that limited the practical advantages of its MI300X hardware. Intel, Google, Microsoft, Meta and others formed an accelerator-interconnect standards group intended to reduce dependence on Nvidia hardware. Triton and other software efforts target CUDA lock-in. OpenAI and Broadcom co-designed an inference chip; Cerebras raised $5.5 billion in its IPO, and Tenstorrent raised $1.5 billion. Nvidia’s lead is substantial, not absolute.

Nor does every route around Nvidia require buying a rival company. Nvidia and Groq entered a confirmed $20 billion non-exclusive agreement licensing Groq’s inference technology, with Groq chief executive Jonathan Ross joining Nvidia. Earlier acquisition speculation was denied. Licensing and hiring can alter access to technology and talent without transferring ownership of the entire firm.

More vendors could give customers leverage, especially if packaging capacity expands outside Taiwan and alternative software matures. They would still need memory, manufacturing slots, power and financing. Competitors may diversify the names on the accelerator purchase order before they reduce the cost of constructing and operating the facility it serves.

The cheapest token still needs a powered rack. Nvidia’s $3.5 billion in guarantees shows who may have to pay if a customer leasing the land beneath it cannot.