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

Wanting a say in who rents the machines you sell is not an argument about price.

Per the Wall Street Journal of August 28, Nvidia has stopped signing new revenue-sharing financing deals with cloud providers.

Two things went wrong: counterparties objected to a clause, and Nvidia's own employees warned it could draw antitrust attention. The clause: chips could go only to customers Nvidia approved.

Announced July 1, 2026, the program gathered $36 billion in commitments in two months.

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

The small cloud firms that rent out AI chips — neoclouds — are short of capital, not customers: ten thousand GB300 chips is a several-hundred-million-dollar project. So they go to a bank.

The bank asks: if the borrower stops paying, what am I left with? Not a building or a ship — a two-year-old GPU whose resale value nobody can predict.

Lenders price that uncertainty by refusing the loan or charging a rate the project cannot carry.

Nvidia's July solution sat in that gap: it pledged to rent idle capacity itself at a fixed hourly rate. No tenant meant Nvidia was the tenant.

The third link matters: the bank now prices not a two-year-old startup but the commitment of a company that reported $96.2 billion of revenue and $59.7 billion of net income in one quarter. Only the guarantor changes.

The cost of borrowed-money investment depends entirely on whose name is on the guarantee.

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

Nvidia earns twice. First on the chip: an ordinary hardware sale at a 75.0% gross margin.

Second on the revenue share: a base hourly rate covers costs, and 50% above it goes to Nvidia.

Base rates were not disclosed; the figures below are round numbers showing how the structure behaves.

Full, the pledge costs nothing; empty, it is a direct cash outflow. Utilization sets the cost.

A party that can cut a guarantee's cost by its own conduct wants control of it.

Strip the clause out and Nvidia holds an unhedged commitment on chip demand; leave it in and the chipmaker decides which AI firms get compute. That tension stalled the program.

What Nvidia Said, and What It Did Not Say

The company did not deny it. Its statement: the July model, which opens compute access to the AI ecosystem, is "still in place and continues to evolve due to high demand."

That defends the model's existence; it does not say whether specific agreements stopped being signed.

Nor is there regulatory action: no Federal Register notice, no investigation, no court ruling. The brake was pulled by Nvidia's own employees and its counterparties, not a government.

A rental program is not a merger but contractual conduct. Mergers require pre-closing notification; conduct claims require a regulator to sue and prove its case — which is why none of this reached an official docket.

Causation on the stock side is hard. Nvidia reported second-quarter results on August 26 and rose 8.74% the next day to $227.98; on August 28 it fell 4.57% to $217.55.

The Nasdaq 100 fund fell 0.68% and the small-cap fund 1.37% that day; the tape was weak and giving back a sharp post-earnings gain is ordinary. No source separates the report's share.

Same Week, Same Company, a Different Legal Form

Two other transactions ran that week; together the three show where the real constraint sits.

On August 21, Nvidia licensed Poolside's model-development software for $6 billion, invested $1 billion and hired 109 engineers — "not an acquisition and it is not an acquihire," Poolside told investors.

Acquisitions above a set size must be reported to U.S. antitrust agencies before closing; the 2026 threshold is $133.9 million. A license plus hires does not qualify.

Three such deals in nine months — roughly $20 billion for Groq in December 2025, $900 million for Enfabrica, $7 billion for Poolside in August — produced no notification against that combined $28 billion.

Senators Elizabeth Warren and Richard Blumenthal asked in March whether the Groq structure bypassed review. No action followed.

The August 27 report is different: per The Information, Nvidia agreed to buy Hugging Face for $12.9 billion. CNBC, TechCrunch and Fortune confirmed; neither company commented.

As a straight acquisition, notification is mandatory and falls in the top fee tier: $2.46 million for transactions above $5.869 billion.

What Is Left

The first two partners were Sharon AI and Firmus: up to 40,000 GB300 chips over six years, up to 170,000 at a 360-megawatt Batam island site. Fate undisclosed.

The $36 billion figure was expected to shrink anyway: as providers find customers, less idle capacity is left. A decline is not evidence the program ended.

What to watch: the reason given for any change in the third-quarter results, and whether the approval clause survives.

There is also a price effect: without the pledge, credit costs for small cloud firms rise, land in their cash flow, and some projects die before they are built. That lowers chip demand — Nvidia's reason for building the program.

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.