NVIDIA Enables Integration of Custom XPUs into NVLink Infrastructure for AI Factories

NVIDIA has unveiled its latest approach to integrating proprietary custom XPUs, developed by hyperscalers and AI-native enterprises, into its established NVLink infrastructure. This integration shifts the focus from individual chip designs to managing scale-up networks, rack-scale architectures, and production-grade software under a unified AI Factory model. Unlike traditional setups where custom accelerators are deployed in isolated silos, the unified model fuses computing resources using NVLink to directly improve operational metrics such as tokens generated per second and energy efficiency per watt. Organizations with custom silicon can leverage this validated ecosystem to deploy large-scale training and inference environments while minimizing time-to-market complexity and engineering overhead. However, maximizing the performance of this unified environment requires comprehensive optimization across the entire platform, including rack designs and supplier ecosystems. Engineering teams must evaluate how their custom silicon specifications align with NVIDIA's full-stack software and networking requirements to resolve potential integration boundaries before deployment.