NVIDIA is moving beyond a simple “sell the chips and wait for the next upgrade cycle” model for part of the AI infrastructure market. In a July 1 announcement, the company introduced a revenue-sharing and credit-support structure that lets AI cloud operators buy NVIDIA infrastructure, sell NVIDIA-powered cloud services, and pass a portion of supported cloud revenue back to NVIDIA.
The model matters because the bottleneck in AI infrastructure is no longer just access to GPUs. It is also access to capital, power, cooling, customer commitments, and enough utilization to make a giant GPU campus pay for itself. NVIDIA is now trying to solve that financing problem in a way that also gives it a recurring, usage-linked stream tied to the capacity it helps bring online.
The first named examples are Sharon AI and Firmus. Sharon AI is planning up to 40,000 NVIDIA Grace Blackwell GB300 GPUs in Australia. Firmus is building a DSX AI factory campus in Batam, Indonesia, expected to scale to 360 megawatts and up to 170,000 NVIDIA GPUs. Together, the projects show how the AI cloud market is starting to look less like ordinary cloud resale and more like project finance attached to an operating platform.
What NVIDIA is changing
Under the new structure, AI cloud providers can procure NVIDIA infrastructure for customers such as AI-native startups, model builders, enterprises, research organizations, and software vendors. NVIDIA still earns product revenue from the infrastructure sale, but it also receives a share of the cloud revenue generated on the supported capacity.
That is a meaningful shift. A hardware vendor usually benefits when a customer buys systems, then benefits again only if the customer returns for a refresh. NVIDIA’s new structure ties part of its upside to how heavily the cloud capacity is used after deployment. The company’s official framing is that emerging AI companies often have demand but not enough conventional financing to secure large-scale compute. The revenue-share model is meant to bridge that gap by aligning NVIDIA, the cloud operator, and end customers around utilization.
For customers, the promise is faster access to large blocks of accelerated computing without waiting through the full sequence of site selection, power procurement, construction, hardware delivery, and bring-up. For NVIDIA, the promise is deeper control over where scarce AI infrastructure lands and a clearer path to recurring revenue from token-scale workloads.
Why Sharon AI and Firmus are useful test cases
Sharon AI’s June 12 announcement gives the smaller of the two initial examples. The company described a six-year compute collaboration with NVIDIA covering 72 megawatts of new data center capacity in Australia, using NVIDIA’s DSX AI factory design and scaling up to 40,000 GB300 GPUs. Sharon AI said its total AI factory capacity would rise to 132 megawatts after the deal, with 102 megawatts contracted to end customers and more than 55,000 NVIDIA GPUs expected by mid-2027.
Firmus is the larger and more geographically interesting example. Its partnership with NVIDIA runs through 2034 and is anchored by a 360-megawatt DSX AI Factory campus in Batam. The company says the agreement covers up to 170,000 NVIDIA accelerators across Grace-Blackwell, Vera-Rubin, and Vera platforms through 2027 and 2028. Firmus also expects between $25 billion and $30 billion from committed offtake agreements during the first six years of the partnership.
Those numbers are not the same as delivered infrastructure. They include “up to” GPU counts, expected capacity, and future customer commitments. Still, they are large enough to show what NVIDIA is trying to normalize: regional AI clouds with hyperscaler-scale ambitions, backed by a supplier that wants a claim on both the hardware sale and the revenue the hardware generates.
DSX makes the financing model more than accounting
The revenue-share plan is closely tied to NVIDIA DSX, the company’s AI factory platform. DSX is not just a brand name for servers. NVIDIA describes it as a stack of reference designs, simulation tools, operations software, cooling and power optimization, and grid-facing controls for designing and running AI factories.
That technical layer is important because a revenue share is only valuable if the facility produces enough sellable compute. NVIDIA’s DSX announcement in May laid out several pieces: DSX reference designs for compute, networking, storage, power, cooling, and controls; DSX Sim for modeling AI factory choices before deployment; DSX OS for lifecycle management, scheduling, health automation, resiliency, and multi-tenant operations; and DSX Flex for adjusting workloads against grid signals.
The most directly financial piece is DSX MaxLPS. NVIDIA says the software and cooling approach can let operators run up to 40% more GPUs at efficient operating points within a fixed power budget, with limited workload impact. In plain terms, NVIDIA is trying to make the key metric “tokens per megawatt,” then sell the stack that improves that metric.
That gives the business model a tighter logic. If power is the binding constraint for AI factories, then every watt has to produce as much useful model output as possible. If NVIDIA helps finance the facility and also supplies the operating stack that governs utilization, it has more leverage over the economics of the whole site.
The risk is circular demand
The obvious concern is that NVIDIA’s customer-financing role could make demand look stronger than it really is. Investors have already watched NVIDIA support or invest in parts of the AI infrastructure ecosystem, and any revenue-share arrangement raises questions about how much of the market is driven by independent end-customer demand versus supplier-assisted expansion.
That does not make the model weak by itself. AI cloud customers genuinely need capacity for training, fine-tuning, agentic inference, and production workloads. Many startups and enterprise AI teams cannot wait years for hyperscaler capacity or build their own facilities. A financing structure that brings regional capacity online faster could be useful.
But the model depends on utilization staying high, customers paying for the capacity, and the hardware retaining economic value long enough to support the contract. GPU generations are moving quickly. If newer systems sharply lower cost per token, older capacity could face pricing pressure before its financing case is fully proven. If AI demand keeps rising, NVIDIA’s revenue-share model could look like a smart way to capture more of the infrastructure value chain. If utilization disappoints, it could look more like vendor financing stretched over very expensive assets.
What to watch next
The most important signal will not be another large GPU number. It will be evidence that these AI factories are filling with paying workloads at sustainable prices. For Sharon AI, that means watching whether contracted capacity turns into operating capacity and whether the company reaches its mid-2027 GPU deployment targets. For Firmus, the key questions are whether the Batam campus comes online on schedule, how its $25 billion to $30 billion of expected committed offtake is structured, and how much demand comes from durable enterprise and AI-native customers rather than short-term compute scarcity.
The broader market should also watch how much NVIDIA discloses about revenue-share economics. The company has explained the structure at a high level, but not the split, duration, utilization thresholds, or accounting treatment that would show how much recurring revenue this model can produce.
For now, the direction is clear. NVIDIA is not just selling the hardware behind AI factories. It is trying to finance them, standardize their design, tune their power economics, and collect a portion of the revenue they generate. That makes the new AI cloud deals less a side program than a preview of how the next phase of AI infrastructure may be paid for.
Sources: NVIDIA’s July 1 AI cloud partnership announcement, NVIDIA’s DSX platform release, Sharon AI’s June 12 compute collaboration announcement, and Firmus Technologies’ Batam AI factory announcement.
