Nvidia’s DSX POD: Building AI Factories in Just 14 Days
The pace of artificial intelligence development is so rapid that traditional data centre construction has become a bottleneck for many companies. To address this challenge, Nvidia has officially announced the DSX POD—a modular infrastructure solution that integrates the company’s latest Vera Rubin computing systems with essential engineering equipment. This system aims to radically reduce the deployment timeline for AI infrastructure from six months to just 14 days.
Factory Precision and Rapid Assembly
The DSX POD concept relies on shifting the construction process to a factory environment. Instead of lengthy on-site assembly, Nvidia’s partners—including Dell Technologies, HPE, Lenovo, Schneider Electric, and Vertiv—manufacture fully equipped containers. Each unit arrives pre-installed with the necessary cabling, electrical wiring, networking infrastructure, and liquid cooling systems.
As Nvidia CEO Jensen Huang noted, "An AI factory creates value when it starts generating tokens, not when construction is finished." This new approach allows businesses to scale their computing power more rapidly and begin training models sooner. However, achieving the 14-day turnaround requires that the site already has the necessary electrical and network infrastructure in place.
Resource Optimization and Environmental Impact
Beyond construction speed, Nvidia is placing a heavy emphasis on energy efficiency. The DSX POD utilizes a closed-loop liquid cooling system that eliminates water evaporation. According to the company’s estimates, in a 2-gigawatt data centre, this approach can save 54 million litres of water per day compared to traditional cooling towers.
Furthermore, the DSX MaxLPS technology dynamically distributes electricity, allowing for the same performance with 30% less energy consumption. Meanwhile, the DSX Flex system enables the data centre to automatically adapt to electricity prices and grid load fluctuations.
Future Outlook
The first DSX POD systems will be based on the Vera Rubin NVL72 architecture, which consists of seven liquid-cooled server racks connected by up to 17 kilometres of Spectrum-X cabling.
While Nvidia’s promises are impressive, their real-world efficiency still requires independent verification. Global shipments are scheduled to begin in early 2027. For countries like Georgia, where data centre infrastructure is actively developing, such "ready-made" solutions could be a significant step toward strengthening the status of a regional AI hub, allowing local companies to adopt international-standard computing power in a short timeframe.