Nvidia Pauses Revenue-Sharing Deals With AI Clouds

Nvidia has paused parts of the revenue-sharing program it unveiled less than two months ago, according to reporting from The Wall Street Journal, retreating from one of the more aggressive financing experiments of the AI buildout. The program, formally called the AI Compute Partnership, was announced on July 1, 2026, and carried commitments totaling roughly $36 billion, with individual deals typically running six years. Some Nvidia employees had reportedly warned customers that the structure could invite antitrust scrutiny, and several prospective partners bristled at the degree of control Nvidia sought over how their capacity was rented out.

The news landed on Friday, August 28, and the market read it as a problem for the companies the program was meant to help. Shares of neocloud providers CoreWeave, Nebius, and IREN all fell, while Nvidia itself slipped in after-hours trading. An Nvidia spokeswoman told the Journal that the underlying July model “is still in place and continues to evolve due to high demand” — a statement that confirms a retreat on the specifics without conceding the strategy.

How the program worked

The AI Compute Partnership was built to solve a real bottleneck. The smaller cloud operators racing to stand up GPU capacity — the so-called neoclouds — need to spend billions of dollars on Nvidia chips and the data centers to house them, but most lack the balance sheet or credit rating to raise that capital on their own terms. Nvidia’s answer was to lend its own financial weight to the purchase.

Under the structure, Nvidia sells the chips, then guarantees to rent back any capacity a customer cannot resell, and finally collects a share of the cloud revenue those chips generate. In other words, Nvidia gets paid at the point of sale and again along the way. For a cash-strapped operator, the backstop transforms the risk calculus: even if end-demand softens, Nvidia’s rent guarantee floors the downside, making it far easier to justify — and finance — an enormous chip order.

That design is why the program mattered well beyond any single deal. It effectively let Nvidia convert its balance-sheet strength into demand for its own products, pulling forward orders that might otherwise have waited on customers’ ability to borrow. The same logic runs through Nvidia’s other recent moves, from a reported $250 billion financing backstop for OpenAI’s Ohio campus to the $500 billion in Wall Street financing it has helped arrange across the ecosystem.

The clause that broke it

Two problems undid the program’s early weeks, and neither was really about the money.

The first was control. According to the reporting, Nvidia pressed partners to rent the subsidized capacity only to approved end users, and to spread it across several smaller customers rather than concentrating it with a single large tenant. For a cloud provider, that is an unusual imposition — it means the chip vendor is helping decide who gets to buy your compute. Several providers pushed back, unwilling to hand a supplier that much say over their own customer relationships.

The second was antitrust exposure. Some Nvidia employees themselves reportedly cautioned current and prospective customers that a chipmaker steering how downstream capacity is sold, while also taking a cut of the resulting revenue, could attract regulatory attention. When a dominant supplier attaches conditions to who its customers may serve, competition authorities tend to take notice — and Nvidia, whose accelerators anchor the AI market, has every reason to avoid handing them a reason to look closer.

Put together, the control provisions and the antitrust worry pointed the same direction. A program meant to grease the flow of chips had instead created friction with the exact partners it was designed to court.

Circular financing, again

The pause revives a debate that has shadowed nearly every large AI-infrastructure headline this year: circularity. Nvidia’s customers need capital to buy Nvidia’s chips; Nvidia supplies or guarantees that capital; the money is spent on Nvidia’s chips, which books as Nvidia’s revenue, which supports the valuation and cash position that let it extend still more financing. The AI Compute Partnership was the most explicit version of that loop yet — Nvidia was not merely selling silicon but underwriting the demand for it and collecting a toll on the output.

The scale of Nvidia’s involvement in its own market is hard to overstate. The company holds more than $40 billion in AI-related equity stakes taken this year alone, and its balance-sheet commitments now touch a long roster of customers and partners. In 2025, it agreed to buy leftover CoreWeave capacity under a deal first valued at roughly $6.3 billion through 2032 — an earlier, one-off version of the backstop logic the AI Compute Partnership tried to systematize.

Defenders note that vendor financing is neither new nor inherently suspect. Aircraft manufacturers, telecom-equipment suppliers, and enterprise-hardware giants have propped up their own demand with credit for decades. The counterargument is about degree and concentration: when the same company supplies the chips, guarantees the rent, takes an equity stake, and shares in the revenue, the risk that would normally sit with a diversified base of lenders instead pools back onto a single balance sheet. Both things can be true at once — the demand can be genuine and the financing structure can still amplify fragility if that demand ever wobbles. This is the same tension running through the broader economics of AI data centers.

Why the neoclouds fell

The market reaction was logical. The AI Compute Partnership was, in practice, a lifeline for operators least able to raise capital independently. Pausing it removes that route for the very companies that needed it most — which is why CoreWeave, Nebius, and IREN took the hardest hits rather than Nvidia.

Neocloud valuations have been unusually sensitive to exactly this kind of news. The group has ridden the AI-compute wave to rich multiples on the premise that demand for rentable GPU capacity is effectively unlimited and that financing will be there to fund the buildout. Nebius, for one, rode that optimism through a strong second-quarter report and a broad neocloud rally earlier in the summer. Anything that complicates the financing side of the story — a paused backstop, a tighter lending market, a supplier stepping back — cuts directly at the growth assumptions baked into those prices.

There is also a strategic read for the neoclouds that competes with the likes of SoftBank’s SB Neo venture and other well-capitalized entrants: without Nvidia’s guarantee, the gap between operators who can self-fund and those who cannot widens. The pause, if it holds, may quietly favor the largest and best-capitalized players over the long tail of smaller providers the program was ostensibly meant to lift.

What it means

The pause is a course correction, not a reversal. Nvidia insists the July model remains in place, and the reporting suggests the company may restructure the initiative or fold it into a different effort later. But the retreat reveals the limits of how far even Nvidia can push its own demand.

Who wins. The best-capitalized cloud operators and hyperscalers, who never needed Nvidia’s backstop, gain relative ground as a subsidized on-ramp for smaller rivals stalls. Nvidia itself avoids near-term antitrust exposure and the reputational cost of being seen to micromanage its customers’ businesses.

Who loses. The smaller neoclouds — CoreWeave, Nebius, IREN, and their peers — lose a financing channel that made aggressive chip orders easier to justify, and their shares repriced accordingly. Any customer mid-negotiation under the program now faces uncertainty about terms.

What to watch next. First, whether Nvidia revives the initiative in a modified form that strips out the customer-approval clauses regulators might question. Second, how the neoclouds fund their next wave of purchases without the rent guarantee — through equity, debt, or tighter coupling to a single hyperscaler tenant. Third, whether antitrust authorities take the reported internal warnings as an invitation to examine Nvidia’s financing arrangements more broadly. And fourth, the read-through to Nvidia’s own numbers: the company posted record quarterly results in late August, but the more of its growth that rests on financing it arranges itself, the more scrutiny each of these structures will draw. For now, the program is paused, not dead — but the episode is a reminder that even the AI cycle’s most powerful company cannot underwrite its own demand without friction.