NVIDIA unveils Scale-In, a new architectural concept that shifts front-end networking from GPU-centric scaling to an accelerated infrastructure fabric that scales users, data, storage, security, control planes, and computing resources together.

Scale-In uses NVIDIA technologies like BlueField-4, evolving from a data processing unit to an AI factory infrastructure processor, and Spectrum-X to support an AI infrastructure network that spans the entire AI factory ecosystem beyond GPU networking.

Scale-In could broaden Spectrum-X’s role from a GPU network to a comprehensive AI infrastructure network, signaling new market opportunities and changes in hardware/software layering.

Agentic AI shifts traffic and workloads beyond model inference to include database queries, storage access, API calls, memory I/O, security validations, and inter-agent communications, pushing bottlenecks outward into the surrounding infrastructure.

Context Scale is proposed as a new memory hierarchy for inference, addressing data locality and memory access as key factors in Agentic AI performance.

NVIDIA Sentry is introduced as a security component for AI agents, extending security considerations beyond the agent’s own server environment.

The article outlines that the AI factory will require five distinct scaling domains, indicating a multi-domain architecture to tackle bottlenecks rather than relying on a single interconnect solution.

The five scaling domains are Scale-Up, Scale-Out, Scale-Across, Scale-In, and a fifth domain related to broader infrastructure orchestration and control planes.