NVIDIA on July 1, 2026 unveiled a new business model that combines revenue-sharing and credit support so AI cloud companies can build large-scale, multi-tenant AI factories. As the first partners, Sharon AI will deploy up to 40,000 NVIDIA Grace Blackwell GB300 GPUs, and Firmus will build a DSX AI factory campus in Batam, Indonesia, scaling to 360 megawatts and up to 170,000 GPUs. NVIDIA earns standard product revenue plus a share of the cloud revenue on the supported capacity, targeting inference providers such as Baseten, Fireworks AI, and Together AI that need fast access to large-scale compute. ASAP summarizes the essentials in a direct-answer format based on NVIDIA's official blog announcement.

From Chip Supplier to Revenue Partner

NVIDIA under this model takes a slice of the cloud revenue on the capacity it supports, on top of the GPU sale itself. According to the July 1, 2026 official blog co-signed by Colette Kress and Raj Mirpuri, the arrangement is designed so AI-native companies, model builders, and inference providers can reach accelerated compute without infrastructure buildout delays.

The core shift is that NVIDIA's position moves from parts supplier to a stakeholder in its customers' income statements. Previously NVIDIA's revenue ended when a GPU shipped; now it keeps earning a share while that GPU runs inside a cloud generating revenue. The center of gravity of the business model has moved from selling one link in the supply chain to sharing in the final service revenue.

How to Read 40,000 GB300s and a 170,000-GPU Campus

Sharon AI will deploy up to 40,000 Grace Blackwell GB300 GPUs, while Firmus will expand its Batam DSX campus to 360 megawatts and up to 170,000 GPUs. NVIDIA connects this capacity to high-demand customers such as Baseten, Fireworks AI, and Together AI for immediate training, post-training, fine-tuning, and inference.

Partner	Scale	Location / Use
Sharon AI	Up to 40,000 GB300	Multi-tenant AI factory
Firmus	360 MW · up to 170,000 GPUs	Batam, Indonesia DSX campus
Target customers	Baseten · Fireworks AI · Together AI	Immediate training/inference

The weight of these numbers sits in the power figure more than the GPU count. 360 megawatts is the continuous load of a mid-sized city, showing that the AI-factory race has moved from chip procurement to a fight over power and land. The 40,000 and 170,000 GPU counts are, in the end, close to a result of how many chips fit inside that power budget.

The Risk of Extending Credit to Your Own Customers

NVIDIA's decision to extend credit to the very companies buying its chips raises circular-financing concerns, with no dollar figure disclosed as of July 1, 2026. The announcement disclosed no specific investment amount or equity percentage, and the settlement mechanics of the revenue share were left unspecified.

There is a point worth stepping back on here. When a chip vendor puts capital into the revenue of the customers who buy those chips, the line between real demand and vendor-financed demand blurs. In the short run compute supply expands quickly, but if the cloud revenue falls short of expectations, the risk NVIDIA has taken on comes back as a boomerang. With the amounts and settlement terms undisclosed, the soundness of this model can only be judged once the first partners' actual utilization rates appear.

What It Means for Korean Cloud and Sovereign AI

For Korea's emerging GPU clouds and sovereign-AI operators, this model becomes a bypass that lowers the upfront capital burden of building capacity. The 360-megawatt campus Firmus is building in Batam, Indonesia, is a regional hub aimed at Southeast Asia, which also signals a shift in the center of gravity of Asia's compute supply chain.

There are two things for Korean practitioners to weigh. One is that the options for reaching large-scale compute through a revenue-sharing contract, without self-financing it, are growing; the other is that in exchange, dependence on the NVIDIA ecosystem deepens all the way into the financial structure. An operator that has secured power and land gains more negotiating cards, but when signing a contract whose settlement terms are undisclosed, it must scrutinize the revenue-share ratio and the utilization assumptions.

---

Source: NVIDIA official blog (announced July 1, 2026, by Colette Kress and Raj Mirpuri)