An AI factory is purpose-built computing infrastructure for the AI lifecycle, from data ingestion and training to inference, according to NVIDIA’s definition. Its network serves several distinct functions rather than acting as one all-purpose connection.
In an interview at the OCP Global Summit, NVIDIA SVP of Networking Gilad Shainer described five roles: scale-up, scale-out, scale-across, context-scale and scale-in. Linda Hardesty reported Shainer’s explanations for Fierce Network on October 14, 2026. The labels describe what parts of an AI factory’s network do; they are not, by themselves, a complete equipment list.
The five network roles in an AI factory
| Network role | Function Shainer described |
| Scale-up | Connects GPUs within a rack-scale system. The reported Vera Rubin NVL72 example uses NVLink and contains 72 Rubin GPUs and 36 Vera CPUs. |
| Scale-out | Connects racks to support larger workloads. Shainer cited training that can involve hundreds of thousands of GPUs. |
| Scale-across | Connects separate data centers so they can operate as one; Shainer described this as useful when one site lacks sufficient power. |
| Context-scale | Provides a separate storage network for inference. |
| Scale-in | Supports secure access for users and security across agents running within the AI factory. |
These roles span different boundaries: GPU-to-GPU communication within a system, links between racks and data centers, storage for inference, and access into the environment. A single AI workload can depend on several of them, but each addresses a different part of the system.
Why distributed GPU work has to stay in sync
Shainer’s explanation starts with a job split among GPUs. They calculate parts of a phase, then aggregate their results before the next phase can proceed. If data reaches the GPUs at different times, the group can be held up while it waits for the required results. That makes coordination and timing part of the networking problem, not just moving data between machines.
In an English-language presentation published by NextGenInfra, Shainer discusses synchronized GPU work and presents diagrams of scale-up, scale-out and scale-across. The presentation gives a visual companion to the roles in the table.
Copper for short links, fiber for longer ones
Shainer described copper as a cost-effective, reliable, low-power choice for short connections, including links within and between racks. He described fiber-optic connections for longer distances. The distinction is about link length; it does not assign a particular distance threshold to either medium.
A role map is not a vendor bill of materials
Cisco describes a Secure AI Factory design with data-center switches based on either Cisco Silicon One or NVIDIA Spectrum-X, alongside separate security, observability, software and storage components. Cisco’s solution description is one vendor’s implementation example. Read alongside Shainer’s functional categories, it illustrates why the five roles do not, on their own, prescribe a single complete vendor stack.
