Nvidia is putting its own balance sheet behind the AI infrastructure buildout, launching a revenue-sharing model that gives AI cloud operators access to large-scale Nvidia hardware in exchange for a cut of the tokens they sell. The first two partners, Sharon AI and Firmus, will deploy up to 40,000 Grace Blackwell GB300 GPUs and 170,000 Nvidia GPUs respectively — north of 210,000 accelerators between them. Firmus is anchoring the deployment in a 360-megawatt campus in Batam, Indonesia.
The mechanism is what matters. Nvidia is calling these deployments DSX AI factories, and the pitch is that AI clouds get capital-efficient access to full-stack Nvidia infrastructure without spending years on site selection, power procurement, construction, and hardware bring-up. In return, Nvidia takes a recurring, usage-linked earnings stream. It is a shift from selling GPUs into a market to co-owning the economics of the compute those GPUs generate.
Nvidia framed the move around a specific observation: AI workloads are shifting from model development to production inference, which means always-on factories manufacturing tokens at scale rather than bursty training runs. That workload profile rewards high utilization and fast time-to-online, both of which reward large multi-tenant deployments over bespoke single-customer clusters. It also rewards operators who do not have to raise conventional debt against speculative demand — which is where Nvidia's credit support comes in.
Key facts
- 01Sharon AI is deploying up to 40,000 Nvidia Grace Blackwell GB300 GPUs under the new DSX AI factory model.
- 02Firmus is building a 360 MW campus in Batam, Indonesia that will scale to 170,000 Nvidia GPUs.
- 03Nvidia is fronting large-scale infrastructure to AI clouds in exchange for revenue-sharing and credit support.
- 04Baseten, Fireworks AI and Together AI are named as inference customers targeting the new capacity.
- 05The program was announced July 1, 2026, ahead of Nvidia GTC Berlin on October 20-22.
Sharon AI's 40,000-GPU commitment is a sovereign-compute play. Cofounder and CEO James Manning positioned the deployment as national-scale AI infrastructure, and the deal structure suggests the company gets to book Nvidia hardware without pre-funding it against long-term customer contracts that traditional lenders have struggled to underwrite.
“This strategic collaboration with NVIDIA marks a pivotal moment in Sharon AI's mission to deliver sovereign, large-scale AI compute infrastructure.”— James Manning, Cofounder and CEO of Sharon AI
Firmus is the bigger of the two builds. The Batam campus will scale to 360 megawatts of power and up to 170,000 Nvidia GPUs, putting it in the range of the largest single-site AI factories publicly announced. Indonesia gives the project cheaper power and land than most US or EU sites, and Batam's proximity to Singapore places it inside a regional network of hyperscale customers.
Firmus co-CEO Tim Rosenfield tied the campus directly to Nvidia's DSX reference design, which standardizes how these factories are built, powered, and networked.
Nvidia named Baseten, Fireworks AI, and Together AI as the type of customer this capacity is aimed at — AI-native inference providers running model training, post-training, fine-tuning, and high-volume agentic workloads. Those companies have historically had to piece together capacity across multiple cloud providers, often waiting quarters for allocations. A dedicated DSX supply designed around their workload profile shortens that queue.
The commercial logic points the same direction as recent moves by Meta and SpaceX, both of which have signaled plans to resell AI compute rather than consume all of it internally. Nvidia is now doing something structurally similar from the hardware side: rather than selling GPUs one hyperscaler at a time, it is standing up a network of aligned AI clouds that all run the same reference stack and share the same economics with Nvidia. This follows the AI-backlash coverage this site ran on local opposition to data-center siting, where 71% of surveyed communities opposed nearby buildouts — placing new capacity in jurisdictions like Batam sidesteps that friction.
The unknowns are the terms. Nvidia has not disclosed the revenue-share percentage, the credit-support structure, or how utilization risk is allocated when demand for a particular AI cloud lags. Analysts covering Nvidia's earnings have flagged that recurring, usage-linked revenue would be a meaningful shift in the company's disclosure model, and neither Sharon AI nor Firmus has broken out the capital contribution split publicly.
For Nvidia, the strategic prize is bigger than the accounting. If DSX AI factories become the default way inference capacity gets built in emerging markets and second-tier clouds, Nvidia locks in its software and networking stack — NVLink, Spectrum-X, CUDA — as the substrate for the next decade of AI services, and captures a share of the tokens sold on top. Details will land at GTC Berlin on October 20-22, and the terms Nvidia discloses there will tell the market whether this is a genuine new business line or a financing wrapper around GPU sales.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
