Nvidia will finance its customers' GPU buildouts in exchange for a cut of their cloud revenue

Nvidia spent the AI boom selling the most sought-after hardware on the planet and letting everyone else worry about how to pay for it. On July 1 it started worrying about that too, and charging for the privilege. In a blog post signed by finance chief Colette Kress and Raj Mirpuri, the company said it will back the GPU buildouts of smaller cloud providers directly, and in exchange take a recurring cut of the cloud revenue those chips go on to earn.

Nvidia's own wording is careful. It calls the arrangement a "revenue-sharing and credit-support model" that lets AI clouds "procure NVIDIA infrastructure ... through economic alignment." Put plainly: a cloud operator buys Nvidia chips, rents access to them, and Nvidia gets paid twice, once on the hardware and again as a share of what the operator collects from customers. Kress describes that second stream as "recurring, usage-linked earnings." That is the kind of revenue Wall Street rewards with a fatter multiple than lumpy hardware sales, which is most of what Nvidia books today.

Why a cloud cannot just buy the chips

The program answers a financing problem that is specific to GPUs, and the blog post never quite names it: residual value. When a young cloud company wants to install 40,000 top-end GPUs, it needs a lender willing to treat that hardware as collateral. Banks are reluctant, because a cluster that costs hundreds of millions today may be worth far less in eighteen months once the next architecture ships. Nvidia's successor to Blackwell, the Vera Rubin platform, is already coming. A lender pricing today's chips as collateral is guessing at a depreciation curve that only Nvidia can see, because only Nvidia decides what the next generation does to the value of the current one.

A buyback commitment fixes that. According to reporting from Tech Times and WinBuzzer on the program's mechanics, Nvidia's support includes guaranteeing to absorb idle capacity, renting or repurchasing GPUs that sit unused at a set price. That sets a floor under the asset, and a floor is what makes it bankable. The lender is now effectively lending against Nvidia's balance sheet rather than against a rack of chips whose future price nobody else can call. It is a credit guarantee that, for now, only Nvidia is positioned to write. The revenue-share percentage itself has not been disclosed, and Nvidia declined to give contract terms beyond the blog post.

The first two names

The two launch partners are already operating at real scale. Sharon AI, an Australian operator listed on the Nasdaq, is deploying up to 40,000 Grace Blackwell GB300 GPUs. Firmus Technologies is building what Nvidia calls a DSX AI factory in Batam, Indonesia, a campus expected to reach 360 megawatts and as many as 170,000 GPUs. Together that is roughly 210,000 chips, a fleet that rivals some mid-tier hyperscalers.

"This strategic collaboration with NVIDIA marks a pivotal moment in Sharon AI's mission to deliver sovereign, large-scale AI compute infrastructure," said cofounder and CEO James Manning. Firmus co-CEO Tim Rosenfield framed it around access: "AI-native companies need access to scalable, energy- and cost-efficient compute infrastructure to compete globally."

For anyone who has not priced one, a GB300 NVL72 is a single liquid-cooled rack that packs 72 Blackwell Ultra GPUs and 36 Grace CPUs and draws about 120 kilowatts, about as much electricity as a small neighborhood, to behave as one machine. Nvidia also pointed to Baseten, Fireworks AI and Together AI as the sort of inference-heavy customers it expects to fill this capacity. The demand it is describing is production inference, the always-on token generation that happens after a model is trained, which is where the compute bills now pile up.

The industry has run this play before

Financing your own customers so they can buy your product has a name, vendor financing, and its most famous outing did not end well. In the late 1990s, telecom equipment makers including Lucent, Nortel and Cisco extended billions in loans to carriers so those carriers could buy their gear. It looked like a growth engine while revenue was climbing. When the financing dried up in 2001, large portions of those loan books went uncollected and the vendors ate the losses.

The parallel is not lost on the market. As Tech Times reported, Wedbush analyst Matthew Bryson wrote that Nvidia's buildouts fit "squarely into the circular investment theme" worrying investors about how durable AI demand really is, while allowing that the same strategy could build a "competitive moat" if Nvidia pulls it off. The tell that separates a flywheel from a loop is simple to state and hard to answer: is genuine outside demand paying for this compute, or is money circling among a small set of connected players who all benefit from the appearance of growth?

Why it may not rhyme

There are honest reasons the telecom comparison is imperfect. Those fiber networks sat almost entirely dark; GPU capacity is being consumed now, at scale, by paying customers. Nvidia's balance sheet is also nothing like Nortel's, so it can absorb some partner failures that would have sunk a mid-tier vendor. And the structure itself is better aligned than a plain loan. A loan can be repaid out of a fresh funding round whether or not the customer's business works. A share of cloud revenue only pays Nvidia if real workloads actually run on the chips, which ties Nvidia's upside to end demand rather than to its own sales.

The open question is the one no filing answers yet. A Bain analysis last September estimated the AI industry needs about $2 trillion in annual revenue by 2030 to justify current spending, and put the sector roughly $800 billion short of that path. Nvidia is now selling the chips, backstopping the chips, and collecting a cut of the output from the same handful of counterparties. Whether that reads as a moat or a warning depends entirely on how much of the resulting revenue comes from customers making real decisions outside the funding loop. Watch two things: whether Nvidia breaks out this revenue in its filings, and whether the sovereign-cloud demand behind Sharon AI and Firmus turns into paying third parties or stays a promise.