Nvidia said on July 6 that it is rolling out a global revenue-sharing model that allows AI startups to secure GPU compute by committing part of their future revenue, rather than paying for large hardware purchases upfront. The company named Sharon AI and Firmus as its first partners, signaling a broader shift from chip sales toward deeper involvement in AI infrastructure operations.
A move beyond one-time GPU sales
For many AI startups, getting enough compute has usually meant either buying Nvidia hardware directly or signing long-term contracts with cloud providers. Both options can strain capital, especially as demand for Blackwell-based GB300 GPUs keeps rising. Under Nvidia’s new structure, cloud partners provide GPU compute services to startups, while Nvidia earns its standard hardware revenue and also takes a share of cloud operating income from those partnerships.
That changes Nvidia’s role. Instead of ending the relationship after the chip sale, the company ties part of its economics to the long-term performance of the infrastructure running on its platform. For fast-growing AI companies with urgent compute needs and limited cash flow, the model opens a path that does not require funding hardware buildouts first.
First deployments center on Sharon AI and Firmus
Sharon AI is deploying as many as 40,000 Nvidia Grace Blackwell GB300 GPUs. Firmus is building its DSX AI factory campus in Batam, Indonesia, with plans to scale to 360 megawatts of power capacity and house up to 170,000 Nvidia GPUs. According to the source material, the Firmus-Batam arrangement had previously been treated as a single-site pilot. Nvidia has now elevated that approach into a formal global commercial strategy.
Sharon AI co-founder and CEO James Manning said the strategic partnership with Nvidia is a key moment in Sharon AI’s plan to build sovereign-scale AI compute. Firmus Technologies co-CEO Tim Rosenfield said AI-native companies need compute infrastructure that is scalable, energy-efficient, and cost-efficient in order to compete globally.
Designed for AI-native companies with heavy compute demand
The report points to companies such as Baseten, Fireworks AI, and Together AI as potential beneficiaries. These firms often need immediate access to cloud compute for model training, post-training fine-tuning, and high-concurrency inference, while their products are still moving from pilot phases into production. In a conventional purchasing or leasing model, they can end up constrained either by capital expenditure or by a shortage of compute at a critical stage.
In Nvidia’s own blog post, the company said AI infrastructure demand is shifting from model development to production inference, with AI factories needing to run continuously and generate tokens at scale. The new structure is intended to help AI clouds adopt Nvidia’s platform faster, while extending its reach beyond major cloud providers to startups, model developers, enterprises, research institutions, and regional AI operators.
The model points to a business strategy built not only on GPU shipments, but also on recurring participation in cloud infrastructure revenue.