NVIDIA’s new push for AI “factories” aims to bring compute access, financing and usage economics under one roof

NVIDIA is trying to solve a bottleneck that has become central to the next phase of artificial intelligence: not just building models, but running them continuously at scale for real users. In a new strategy, the company is partnering with AI cloud companies to package large-scale, multi-tenant accelerated compute into “AI factories” that can be brought online quickly and kept highly utilized, aligning the economics of production inference with how much customers actually use.

The push reflects an industry shift from training, which is often done in batches, to inference at production scale, where systems generate outputs continuously and must be ready to serve demand that can spike. NVIDIA’s framing is that compute demand is moving toward token-scale services, meaning the business value is increasingly tied to how efficiently companies can generate tokens, the basic units of text or other generated outputs produced by AI models.

According to NVIDIA, many emerging AI companies have struggled to obtain capital-intensive infrastructure even when they can secure long-term commitments. The company says the gap is financing itself, not just interest from customers, arguing that traditional approaches can leave compute access slow or constrained until facilities, power and hardware bring-up are complete.

To address that, NVIDIA says it will “open up compute access” to a broader AI ecosystem that includes startups, model builders, enterprises, research organizations and regional AI players. The intent is to give AI clouds access to large-scale NVIDIA infrastructure while using a model that combines revenue sharing with credit support, so that capacity expansion can happen without requiring each operator to shoulder the full upfront investment burden.

Under the approach described by NVIDIA, AI cloud companies will sell cloud services delivered through NVIDIA DSX AI factories, which NVIDIA characterizes as systems that manufacture tokens at scale. NVIDIA says the structure is meant to accelerate adoption of NVIDIA platforms among cloud customers, while giving NVIDIA a new recurring revenue stream that is linked to usage.

NVIDIA provided early examples of the strategy taking shape. Sharon AI, a company focused on sovereign AI compute infrastructure, says it is deploying up to 40,000 NVIDIA Grace Blackwell GB300 GPUs as part of the collaboration, positioning the rollout as a way to deliver large-scale AI capacity under local control.

Firmus Technologies, another early partner, says it is building a DSX AI factory campus in Batam, Indonesia. Firmus expects the campus to scale to 360 megawatts and up to 170,000 NVIDIA GPUs, describing the project as an energy- and cost-efficient platform intended to let its AI cloud serve more customers as demand grows.

The company also points to what it calls a new commercial reality for AI-native businesses. It says organizations like Baseten, Fireworks AI and Together AI illustrate where compute demand is heading, as they work through model development steps including training, post-training and fine-tuning, alongside high-volume agentic inference, which refers to AI systems that can take actions as well as generate responses. Their customers, NVIDIA adds, need reliable access to accelerated compute as usage scales, but they also need flexibility as products move from pilots to production.

NVIDIA’s described model is also aimed at shortening timelines that typically slow down AI deployments. For inference providers, agent platforms and enterprises scaling AI, NVIDIA says the partnership can provide full-stack accelerated computing faster than waiting through site selection, power procurement, construction and hardware bring-up.

What NVIDIA does not provide in the announcement is the detailed financial structure of the revenue-sharing and credit-support terms, including whether specific customers will have minimum purchase commitments or how long credit backing lasts. The company also does not disclose pricing, capacity reservation mechanisms, or the timeline for broader availability of the “cloud partner” program beyond the stated early participants. For operators in this space, those specifics may determine whether the arrangement meaningfully reduces capital risk or simply reshapes it.