TL;DR
HPCwire disclosed the first engineering parameters of NVIDIA's DSX POD on October 23, 2026: one container integrates 7 Vera Rubin NVL72 racks with about 10 million components, a total weight of 28,000 kg, and 3,500 cables, with racks, cabling, and liquid cooling assembled and tested in the factory.
Background: The Vera Rubin Platform Enters Volume Production
The NVIDIA Vera Rubin platform is a multi-rack POD-scale system built for agentic AI. Per NVIDIA's official specification, a single Vera Rubin NVL72 rack integrates 72 Rubin GPUs, 36 Vera CPUs, ConnectX-9 SuperNICs, and BlueField-4 DPUs, interconnected in-rack through sixth-generation NVLink and scaled out with Quantum-X800 InfiniBand and Spectrum-X Ethernet. An NVIDIA page updated on October 20, 2026 states that Vera Rubin has ramped into full production, with Taiwan's top server makers and the global supply chain manufacturing and shipping Vera Rubin-based systems at scale.
The Disclosure: First DSX POD Engineering Parameters
According to a Baidu AI (BAAI) Hub summary of HPCwire coverage published on October 24, 2026, HPCwire's October 23 report provided the first engineering parameters of the DSX POD:
| Parameter | Value |
|---|---|
| Vera Rubin NVL72 racks per container | 7 |
| Total components | about 10 million |
| Total weight | about 28,000 kg |
| Total cables | about 3,500 |
| Copper cabling length | about 20 km |
| Scale-out cabling length | about 17 km |
| Assembly method | racks, cabling, and liquid cooling completed and tested in the factory |
Engineering Significance: Factory Pre-Integration and Rapid Deployment
DSX POD moves rack assembly, copper cabling, and liquid-cooling connections from the data center site into the factory environment. This model shortens on-site construction time and reduces manual error, turning a single container into a transportable, ready-to-run compute unit. For AI factory operators expanding under fixed power and space budgets, containerization also adds site-location flexibility and the potential for repeatable bulk deployment.
Industry Context: Standardizing AI Factory Infrastructure
NVIDIA's Vera Rubin NVL72 MGX racks, introduced at the OCP Global Summit, are designed around 45-degree-Celsius warm-water liquid cooling, fully liquid-cooled compute trays, and liquid-cooled busbars rated up to 5,000 A. NVIDIA's technical blog states that 45-degree-C inlet operation lets facilities rely more often on free cooling, freeing enough power to allocate up to 10% additional NVL72 racks within the same power budget. The containerized DSX POD extends this "whole-system delivery, liquid-cooling-first" approach, marking a shift from custom-built data center rooms toward standardized AI factory units.
Sources
- BAAI Hub summary of HPCwire (republished October 24, 2026; original report October 23, 2026): https://hub.baai.ac.cn
- NVIDIA Vera Rubin official page: https://www.nvidia.com/en-in/data-center/technologies/rubin/
- NVIDIA technical blog (MGX architecture and 45-degree-C liquid cooling): https://developer.nvidia.com/blog/?p=113993
- NVIDIA blog (Vera Rubin NVL72 MGX and AI factories): https://blogs.nvidia.com/blog/gigawatt-ai-factories-ocp-vera-rubin/