Definition

An AI factory is a large-scale data center designed specifically for training and running frontier AI models, typically requiring hundreds of thousands of GPUs connected via high-performance networking infrastructure.

Key Technical Requirements

High-Performance Networking

AI factories require specialized networking to handle:

  • Elephant flows: Large data transfers that can saturate single network paths
  • Tail latency: Delays that hinder training efficiency
  • Massive scale: 100,000+ GPUs in a single cluster

NVIDIA MRC Protocol

The Multipath Reliable Connection (MRC) protocol addresses these requirements through:

  • Packet spraying: Distributing packets across all available network paths
  • Microsecond-level rerouting: Dynamic steering around congestion
  • Connection pooling: Scaling connection management per node
  • Hardware-accelerated load balancing: CPU-free failure detection

Deployment Examples

  • Japan noetra Vera Rubin factory (announced Jul 2026): ~27,500 Rubin GPUs + 13,750 Vera CPUs, ~140 MW, Spectrum-X/BlueField; construction Apr 2027, ops Jun 2028 — compute for frontia-project (2026-07-16-nvidia-japan-noetra-national-ai-infrastructure)
  • Microsoft Fairwater: AI factory infrastructure
  • Oracle Cloud Infrastructure Abilene: AI data centers
  • OpenAI frontier model clusters: Training infrastructure

Key Points

Sources