Nvidia lets AI startups pay later for compute now

Nvidia is shifting its financial strategy to help artificial intelligence startups overcome a major barrier: the high upfront cost of computing power. Rather than simply selling chips for immediate payment, the company has introduced a revenue-sharing and credit-support model designed to get its GPUs into the hands of AI cloud providers that might otherwise struggle to afford them.

Announced on Wednesday, this new approach allows AI cloud companies to access large quantities of Nvidia’s hardware in exchange for a portion of the future revenue those chips generate. Nvidia frames this as a solution to a capital bottleneck that has long hindered emerging AI firms. These startups often lack the resources to invest in the infrastructure required to train and run large models, and even long-term customer commitments have not always been enough to secure financing for compute resources.

Under the arrangement, AI clouds can purchase Nvidia’s chips and resell the computing power, with Nvidia collecting standard product revenue upfront and then an additional share of whatever the cloud earns from renting out the GPUs. This model mirrors the dynamics that have driven up valuations for GPU resellers like Runpod, which reached a $1 billion valuation in June by renting chips it does not own.

Two companies have already signed on. Sharon AI, an Australian AI cloud operator, is deploying up to 40,000 Nvidia Grace Blackwell GB300 GPUs under a six-year, 72-megawatt agreement. Its cofounder and CEO, James Manning, called the deal “a pivotal moment” for the company’s push into sovereign, large-scale AI compute. Firmus, the other early partner, is developing a much larger campus: a 360-megawatt Nvidia DSX AI factory in Batam, Indonesia, that will eventually house up to 170,000 GPUs across Nvidia’s Grace-Blackwell, Vera-Rubin, and Vera platforms. Bloomberg has reported that Firmus anticipates between $25 billion and $30 billion in committed offtake agreements over the first six years, a scale that assumes continued demand from AI-native customers. Nvidia named Baseten, Fireworks AI, and Together AI as examples of the customers this model is intended to serve.

These clients need immediate, elastic access to AI cloud capacity for training, fine-tuning, and high-volume inference without committing to years of hardware procurement themselves. They represent a different customer base from the hyperscalers Nvidia has courted for the past decade. The model is a bet on the long tail of model builders, agent platforms, and enterprises that want frontier compute but not the balance-sheet risk of constructing a data centre.

For Nvidia, this arrangement creates something new: a recurring, usage-linked income stream layered on top of hardware sales. The model pairs revenue sharing with credit support, effectively helping smaller AI clouds finance their purchases. It is not a loan, but it functions like vendor financing with an equity-like upside. The chips themselves still cost what they cost, and Nvidia is not changing what it sells. What changes is who can afford to buy them and on what terms.

This shift matters because site selection, power procurement, construction, and hardware bring-up can take years before a startup ever runs a workload. Nvidia’s pitch is that AI cloud partners can compress that timeline by selling capacity that already exists. The company has already committed more than $40 billion to direct AI equity investments this year, spanning OpenAI, Nebius, and dozens of smaller rounds. A revenue-sharing compute model does something similar without touching the cap table, keeping balance-sheet exposure with its cloud partners rather than on its own books.

Nvidia has not disclosed how many AI clouds it expects to sign on this basis, or whether the terms for Sharon AI and Firmus will be standardised across future partners. The model also deepens a dependency that has already drawn scrutiny, as an increasing share of the AI industry’s growth becomes contractually tied to Nvidia’s success. If the model works, more compute reaches more startups faster than the traditional buy-it-outright approach allowed. If AI-native demand cools, Nvidia is now exposed to that slowdown twice: once through chip sales and again through the cloud revenue it has agreed to share.