====== AI Superfactory ======

An **AI Superfactory** (also called an **AI Factory**) is NVIDIA's term for a purpose-built computing facility designed specifically to manufacture artificial intelligence — training models, running inference, and generating tokens at industrial scale. Unlike traditional data centers optimized for storing and retrieving data, AI Superfactories are engineered for continuous, high-density compute workloads with specialized power, cooling, and networking infrastructure. ((Source: [[https://nvidianews.nvidia.com/news/blackwell-ultra-dgx-superpod-supercomputer-ai-factories|NVIDIA — Blackwell Ultra DGX SuperPOD]]))

===== Concept =====

Jensen Huang, NVIDIA's CEO, has described AI factories as the defining infrastructure of the AI industrial revolution — facilities that take in raw data and produce intelligence as their output, analogous to how traditional factories take in raw materials and produce physical goods. ((Source: [[https://nvidianews.nvidia.com/news/nvidia-partners-ai-infrastructure-america|NVIDIA — AI Infrastructure America]]))

Key distinctions from traditional data centers:

  * **Workload**: Continuous GPU-intensive AI training and inference vs intermittent general-purpose computing
  * **Power density**: 50-100+ kW per rack vs 5-15 kW for traditional servers
  * **Cooling**: 100% liquid cooling at 45°C inlet vs air-cooled environments
  * **Networking**: NVLink fabric for GPU-to-GPU communication at hundreds of TB/s vs standard Ethernet
  * **Scale**: Approaching gigawatt power consumption per campus

===== Key Components =====

==== DGX SuperPOD ====

The **DGX SuperPOD** is NVIDIA's reference architecture for AI factories, built from modular **Scalable Units (SUs)**:

  * Each SU contains 8 DGX systems for rapid deployment
  * Current generation uses **DGX GB300** systems with Grace CPUs and Blackwell Ultra GPUs
  * Next generation will use Vera Rubin-based systems ((Source: [[https://docs.nvidia.com/pdf/dgx-spod-gb300-ra.pdf|NVIDIA — DGX SuperPOD GB300 Reference Architecture]]))

==== Networking ====

  * **NVIDIA Quantum-X800** InfiniBand (XDR/800 Gbps) for high-performance, low-latency interconnect
  * **NVLink** fabric for scale-up GPU-to-GPU communication within racks
  * **Spectrum-X** Ethernet for scale-out connectivity
  * **BlueField DPUs** for infrastructure offload and security

==== DSX Reference Design ====

The **Vera Rubin DSX AI Factory** reference design (announced GTC 2026) provides a codesigned infrastructure guide for maximum tokens-per-watt and accelerated time to first production. It pairs with the Omniverse DSX Blueprint for digital twin simulation of the entire facility. ((Source: [[http://nvidianews.nvidia.com/news/nvidia-releases-vera-rubin-dsx-ai-factory-reference-design-and-omniverse-dsx-digital-twin-blueprint-with-broad-industry-support|NVIDIA — Vera Rubin DSX]]))

===== Who Is Building AI Superfactories =====

  * **NVIDIA** — AI Factory Research Center in Virginia hosting first Vera Rubin infrastructure
  * **U.S. Department of Energy** — Seven systems across Argonne and Los Alamos National Laboratories
  * **Cloud providers** — AWS, Google Cloud, Microsoft Azure, Oracle Cloud, CoreWeave
  * **Data center operators** — Switch (EVO AI Factories), Equinix (Instant AI Factory service)
  * **Enterprises** across server makers, model builders, and technology suppliers are investing in dedicated AI factory infrastructure ((Source: [[https://nvidianews.nvidia.com/news/nvidia-partners-ai-infrastructure-america|NVIDIA — AI Infrastructure America]]))

===== Scale =====

The DGX SuperPOD architecture is designed as a "physical twin" of NVIDIA's own internal R&D systems, ensuring that software, applications, and support are pre-tested on identical infrastructure. Using Scalable Units, deployment times are reduced from months to weeks. ((Source: [[https://docs.nvidia.com/dgx-superpod/reference-architecture/scalable-infrastructure-b300/latest/dgx-superpod-components.html|NVIDIA — DGX SuperPOD Components]]))

===== See Also =====

  * [[nvidia_vera_rubin|Nvidia Vera Rubin]]
  * [[nvidia_omniverse_digital_twins|Nvidia Omniverse Digital Twins]]
  * [[gigawatt_ceiling|The Gigawatt Ceiling]]

===== References =====