Nvidia Hits Pause on Its $36B AI Cloud Financing Program Amid Antitrust Fears

On July 1, 2026, Nvidia unveiled what looked like a clever fix to the AI buildout’s oldest problem: smaller GPU cloud providers — the so-called “neoclouds” — struggle to finance billion-dollar data centers because lenders cannot be sure the capacity will ever be rented. Nvidia’s answer was the AI Compute Partnership, a program that offered neoclouds credit support and demand commitments in exchange for a share of the revenue their GPUs generated. Less than two months later, the Wall Street Journal reports that Nvidia has paused some of those deals — and the reason is the one thing the company cannot afford: antitrust exposure.

What the program actually did

To understand why the pause matters, you have to understand how unusual the arrangement was. Nvidia does not lend money or directly finance data center construction. Instead, it provides something lenders value almost as much: a take-or-pay commitment on a portion of a new facility’s capacity, plus a minimum revenue guarantee, typically running six years. That guarantee gives banks the confidence to underwrite projects that would otherwise look like speculative bets on future AI demand.

“Nvidia provides a take-or-pay commitment on a portion of the facility’s capacity, a minimum revenue guarantee that gives lenders the confidence to underwrite the project, and in exchange, we share in a portion of the NeoCloud’s revenue earned above that floor,” CFO Colette Kress explained on the company’s earnings call. “In this model, we get paid twice: once on the hardware sale, and again through the share of rental revenue, a highly recurring stream layered on top of a one-time equipment purchase.”

The economics work on a simple floor-and-upside logic. If a facility rents strongly and revenue soars past the guaranteed minimum, Nvidia collects a percentage of the excess — by the WSJ’s account, 50% of revenue above a base hourly-rate threshold. If demand is weak and the facility sits half-empty, Nvidia’s take-or-pay obligation forces it to cover the gap or rent the unused capacity for itself. Either way, the neocloud gets financing it could not obtain alone, and Nvidia converts a one-time chip sale into a recurring income stream.

The first two partners announced in July set the scale of the ambition. Sharon AI committed to deploying up to 40,000 Grace Blackwell GB300 GPUs, while Firmus Technologies is building a 360-megawatt AI factory campus in Batam, Indonesia, designed to host up to 170,000 GPUs — together north of 200,000 GPUs of potential capacity. And the program grew fast: in its quarterly filing with the SEC, Nvidia disclosed that its commitments under the program — “typically six years in duration” — had already reached $36 billion as of July 26, 2026, just weeks after launch.

Why it stalled

According to the WSJ report, two forces collided to force the pause. The first came from inside the company: Nvidia employees warned existing and prospective customers that the arrangement “could invite antitrust scrutiny,” given how much control Nvidia would gain over companies that are simultaneously its customers, its financiers’ beneficiaries, and its competitors’ potential rivals.

The specifics of that control are what made lawyers nervous. Nvidia reportedly told some participating cloud providers that they could lease its GPUs only to customers approved by Nvidia. It also preferred to spread available capacity across multiple smaller AI companies rather than let a single large customer absorb most of it — reasonable utilization logic on its face, but indistinguishable from a dominant supplier dictating who may buy from whom downstream.

The second force was the partners themselves. Cloud operators pushed back against the restrictions, arguing they should retain control over which customers they serve. Some potential partners walked away frustrated during the program’s early stages. The friction was predictable in hindsight: neoclouds exist because the market wanted GPU capacity free of hyperscaler gatekeeping, and Nvidia’s terms threatened to recreate exactly that gatekeeping one level down.

Nvidia, for its part, denies that anything fundamental has changed. “The new business model we introduced in July that opens up compute access to the fast-growing AI ecosystem is still in place and continues to evolve due to high demand,” a spokesperson told Tom’s Hardware. That word “evolve” is doing heavy lifting — the report does not claim the program is dead, only that some transactions are on hold while the terms are, presumably, restructured.

The antitrust shadow

The uncomfortable context here is that Nvidia has spent 2026 steadily expanding from chipmaker into something closer to a vertically integrated AI infrastructure financier. It has backed $500 billion AI infrastructure funds with major financial institutions, taken stakes in neoclouds, and now guaranteed revenue for the very customers who buy its silicon. Each step is individually defensible as “meeting customers where they are.” Collectively, they describe a company whose chips, capital, and customer lists sit at the center of nearly every major AI buildout on earth.

That position invites exactly the scrutiny employees feared. A supplier that controls 90%-plus of AI accelerator supply, then attaches conditions on which downstream customers its financed clouds may serve, starts to look less like a vendor and more like an orchestrator of the market — the kind of arrangement antitrust regulators in Washington and Brussels have spent the past two years studying. The customer-approval clause is the flashpoint: whatever its operational rationale, “you may only rent to whom we approve” is the sentence most likely to appear in a future complaint.

There is also the circularity critique. Nvidia does not lend money directly, but by guaranteeing demand for facilities built around its own hardware, collecting 50% of upside revenue, and in weak-demand scenarios renting the capacity back for itself, it tightens a loop in which Nvidia’s capital effectively underwrites purchases of Nvidia’s products. Investors have already been probing the broader circular-financing pattern across the AI ecosystem this year; the pause suggests Nvidia’s own staff concluded the optics had crossed a line.

What happens next

The most likely outcome is not abandonment but restructure. The core take-or-pay mechanism remains attractive to everyone: neoclouds get cheaper capital, lenders get an investment-grade anchor tenant, and Nvidia gets recurring revenue on top of hardware sales. The problematic parts — the customer-approval rights and capacity-allocation preferences — are the terms most likely to be quietly softened or dropped as the program “evolves.”

For neoclouds, the episode is a cautionary tale about the price of cheap capital. Financing that comes with strings attached to your customer list is not really financing; it is a governance transfer. The providers that pushed back understood that accepting Nvidia-approved customer rosters today would make them permanently dependent on Nvidia’s goodwill — and permanently less interesting to any large customer who resents a chipmaker standing between them and their cloud.

For the broader AI infrastructure market, the pause is a reminder that the industry’s most acute tension is no longer supply versus demand but independence versus dependence. Everyone building AI infrastructure needs Nvidia. The question this episode forced into the open is how much of their autonomy they must surrender to get it — and whether the dominant supplier of the AI era can resist using financing as leverage. Nvidia’s employees, it turns out, asked that question before the regulators did. Pausing the program may cost the company some momentum in the short term. Continuing it unchanged might have cost far more.