NVIDIA Unveils Scale-In: Revolutionizing AI Infrastructure Beyond GPU-Centric Networking
September 30, 2026
Nvidia is building an AI agent security stack to govern what autonomous systems can access and do across files, code, and APIs.
Context Scale is proposed as a memory hierarchy for inference, aiming to optimize data locality and memory access in Agentic AI workloads.
Disclaimer: HokaNews content is informational and not financial advice.
The security approach relies on infrastructure-level safeguards that operate independently of the AI model to enforce permissions.
Agentic AI shifts workloads beyond model inference to include database queries, storage access, API calls, memory reads/writes, security validations, and inter-agent communication, moving bottlenecks from GPUs to the broader infrastructure.
The AI factory design calls for five distinct scaling domains to tackle different bottlenecks, signaling a multi-domain architecture.
Five scaling domains are Scale-Up, Scale-Out, Scale-Across, Scale-In, and an inferred domain covering infrastructure orchestration and control planes.
Coin Bureau describes Nvidia’s approach and highlights treating AI security as an infrastructure problem.
Agent-based AI requires stricter permission controls since agents can directly interact with external tools and systems.
A layer between the AI system and resources will govern which actions agents are permitted to perform.
Scale-In is introduced to redefine front-end networking for Agentic AI, expanding the focus from GPU-centric scaling to an accelerated infrastructure fabric across users, data, storage, security, and control planes.
Scale-In builds on NVIDIA tech like BlueField-4 and Spectrum-X to support a broader AI infrastructure network beyond GPU networking.
Summary based on 2 sources
Get a daily email with more AI stories
Sources

SEMIVISION @_@ • Sep 30, 2026
From Scale-Up and Scale-Out to Scale-In: How Agentic AI Is Redefining NVIDIA’s AI Factory Network Architecture