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
- 2026-08: valar-atomics + nvidia plan waterless 30MW nuclear-powered AI factory (Utah) post-$1B Series B (2026-08-03-valar-atomics-tnw-series-b)