Nvidia Paused Its Chip-Rental Program: It Wanted to Pick the Renters

When a company wants a say in who rents the machines it sells, the argument is no longer about price. Price is not what stopped the program Nvidia built in July, two months after it started.

According to the Wall Street Journal's August 28 report, Nvidia has stopped signing new revenue-sharing financing agreements with cloud providers. Two things went wrong at once: the counterparties objected to a clause in the contract, and Nvidia's own employees warned that the same clause could draw antitrust attention. The clause said the chips could be rented only to customers Nvidia approved.

The program was announced on July 1, 2026. It gathered $36 billion in commitments in two months. What follows is why that clause was written into the contract, and why taking it out is not simple.

The Lender's Problem: A Used Chip Has No Price

The small cloud companies that rent out AI chips — the industry calls them neoclouds — share one difficulty. They are not short of customers. They are short of capital. Installing ten thousand GB300 chips is a several-hundred-million-dollar project, and most of these firms do not have the money. So they go to a bank.

The bank asks a simple question: if the borrower stops paying, what am I left with? A factory leaves a building. A ship leaves a ship. Here what is left is a two-year-old graphics processor whose resale value nobody can predict. Lenders price that uncertainty either by refusing the loan or by charging a rate the project cannot carry.

The solution Nvidia proposed in July sat exactly in that gap. The company pledged to rent any idle capacity itself, at a fixed hourly rate agreed in advance. If the provider could not find a tenant, the tenant would be Nvidia.

The third link is the one that matters. The pledge changes who is being underwritten: the bank is no longer pricing the creditworthiness of a two-year-old cloud startup but the commitment of a company that reported $96.2 billion in revenue and $59.7 billion in net income in a single quarter. Same project, same chips, same customers — only the guarantor changes, and the loan becomes possible. The cost of an investment made with borrowed money depends entirely on whose name is on the guarantee.

The Mechanism: When the Pledge Is Free and When It Is Not

Nvidia earns twice here. First on the chip itself, an ordinary hardware sale carried at a 75.0% gross margin last quarter. Second on the revenue share: per the Journal, the two sides set a base hourly rate covering costs, and 50% of everything above that base goes to Nvidia.

The base rates in the actual contracts were not disclosed. The figures below do not come from a real agreement; they are round numbers chosen to show how the structure behaves.

The gap between those two outcomes explains why the program was built this way. While the machines are full, the pledge costs Nvidia nothing. While they are empty, it is a direct cash outflow. Nvidia has, in effect, written a commitment on demand for its own chips, and the cost of that commitment is set entirely by the utilization rate.

When a party can reduce the cost of a guarantee through its own conduct, it will want control of that conduct. If Nvidia picks the tenant, the vacancy risk largely disappears: it steers demand from its own network into the site it has guaranteed, and the pledge never comes due. The second preference the Journal reported follows the same logic. Nvidia wanted capacity spread across several smaller AI firms rather than leased to one large operator. In a single-tenant site, one departure empties the building; in a five-tenant site, one departure empties a fifth of it. Spreading the tenants lowers the odds of the pledge being triggered.

The approved-customer clause was therefore not decoration. It was the pricing leg of the guarantee. Strip it out and Nvidia is left holding an unhedged commitment on chip demand. Leave it in and the company that makes the chips decides which AI firms get access to compute. That tension is what stalled the program.

What Nvidia Said, and What It Did Not Say

The company did not deny the report outright. Its statement was that the business model introduced in July, which opens up compute access to the AI ecosystem, is "still in place and continues to evolve due to high demand."

That sentence defends the existence of the model. It does not answer whether specific agreements have stopped being signed. Those are different claims. A model being in place does not mean new deals are still being written under it.

Nor is there any regulatory action here. There is no Federal Register notice, no opened investigation, no court ruling, no consent decree. The brake was pulled not by a government but, per the Journal, by Nvidia's own employees and by the companies on the other side of the table. The distinction matters: a rental program is not a merger, it is contractual conduct. Mergers carry a mandatory pre-closing notification; conduct claims require a regulator to file suit and prove its case over a period of years. That is why none of this appeared on any official docket.

Causation on the stock side is hard to establish. Nvidia reported second-quarter results on the evening of August 26 and rose 8.74% the next day, to $227.98. On August 28, when the Journal's report ran, it fell 4.57% to $217.55. The Nasdaq 100 fund was down 0.68% and the small-cap fund down 1.37% the same day, so the tape was broadly weak, and giving back part of a sharp post-earnings gain the following session is ordinary behavior. No source separates out how much of the decline belongs to the report.

Same Week, Same Company, a Different Legal Form

Two other Nvidia transactions were in play the week the program stalled, and reading all three together shows where the real constraint on the company's expansion now sits.

On August 21, Nvidia licensed Poolside's model-development software for $6 billion, invested $1 billion in the company and hired 109 of its engineers. The letter Poolside sent its investors described the transaction as "not an acquisition and it is not an acquihire."

That description carries legal consequences. In the United States, acquisitions above a set size must be reported to the antitrust agencies before closing; for 2026 that threshold is $133.9 million. A license plus a hiring wave does not meet the definition. Nvidia has done three deals of this shape in nine months — roughly $20 billion for Groq in December 2025, roughly $900 million for Enfabrica, and $7 billion for Poolside in August — and not one notification was filed against that combined $28 billion. Senators Elizabeth Warren and Richard Blumenthal wrote to the company in March asking whether the Groq structure had been built to bypass the review process. No action has followed.

The report that emerged on the evening of August 27 is a different animal. According to The Information, Nvidia has agreed to buy Hugging Face, the platform that hosts open-source AI models, for $12.9 billion; CNBC, TechCrunch and Fortune confirmed it the same day, and neither company has commented. Because that is a straight acquisition, notification is mandatory and it falls in the top fee tier, $2.46 million for transactions above $5.869 billion. So $28 billion of licensing produces no filing while a single $12.9 billion purchase produces one. What sets the threshold is not the size of the money but the form of the transaction.

What Is Left

The program's first two partners were Sharon AI and Firmus Technologies. Sharon AI planned up to 40,000 GB300 chips under a six-year agreement; Firmus planned up to 170,000 chips at a 360-megawatt site on Batam island in Indonesia. The fate of those contracts has not been disclosed.

The $36 billion commitment figure was expected to shrink on its own: as providers find their own customers, there is less idle capacity for Nvidia to rent. A decline in that number is therefore not by itself evidence the program has ended. What to watch is the reason given for any change in the third-quarter results, and whether the approval clause survives in the contract language.

There is also a price effect. If the pledge goes away, the cost of credit for small cloud firms rises, that cost lands in their cash flow, and some projects are cancelled before they are built. That in turn lowers demand for the chips. Which is why Nvidia built the program in the first place.

This piece is based on the Wall Street Journal's August 28, 2026 report and the secondary accounts of it published by Seeking Alpha and Benzinga, on Nvidia's second-quarter results as filed with the SEC, on the Federal Trade Commission's 2026 threshold announcement, and on the letter text published by the U.S. Senate. The details of the rental program's suspension rest on unnamed sources; Nvidia has not confirmed the report directly and has stated that the model remains in place. Neither party has commented officially on the Hugging Face transaction. This is not investment advice.