NVIDIA Introduces BlueField-4 and Scale-In Infrastructure for Agentic AI

NVIDIA is introducing Scale-In network infrastructure as the fifth pillar of its AI networking. [1]

Powered by the BlueField-4 DPU and NVIDIA DOCA software, this architecture connects over Spectrum-X Ethernet. [1] This approach evolves north-south networks into a coordinated infrastructure domain specifically designed for the AI factory. [1]

Agentic AI factories require multi-terabit bandwidth per server to connect diverse users, agents, and data sources to massive compute. [1]

By keeping infrastructure services off host CPUs, BlueField-4 provides dedicated processing that prevents security and operations from becoming bottlenecks. [1] Consequently, this creates a more efficient and secure shared AI infrastructure as demand scales. [1]

NVIDIA is clearly delineating the networking stack into distinct operational domains, separating "Scale-Up" for GPU coherence via NVLink from "Scale-In" for north-south infrastructure management. By introducing BlueField-4 alongside Spectrum-X Ethernet, the company aims to ensure that peripheral security and data routing operations do not drag down expensive core GPU compute cycles in highly demanding, multi-terabit agentic workflows. What the sources don't address: How the pricing and enterprise availability of BlueField-4 will impact the overall cost of deploying these advanced agentic AI factory architectures.