During the GTC Berlin conference, Nvidia unveiled a new reference design known as DSX POD, aimed at transforming the way artificial intelligence data centers are built by pre-configuring them inside shipping containers before they arrive at the final site. This innovation relies on assembling and connecting racks, networking, and cooling systems in factories to reduce actual operational deployment time.
Traditionally, installation and cabling operations for racks inside conventional data centers take many long months. According to the company's statements, using the DSX POD design can make the system ready to produce tokens in just 14 days, provided that the basic infrastructure for power, cooling, and network connectivity is available on-site.
DSX POD Container Specifications
Each container unit contains seven Vera Rubin NVL72 racks, which collectively house 504 Rubin Graphics Processing Units (GPUs). According to Ian Buck, Vice President of Hyperscale and High-Performance Computing at Nvidia, a single unit consists of about 10 million components and weighs nearly 28 metric units (tonnes).
In addition, the containers come pre-equipped with about 20 kilometers of copper cabling and 17 kilometers of pre-tested expansion networks. Once the container arrives on-site, work crews install it, connect it to adjacent units, and perform final burn-in tests.
Dry Cooling Design to Save Water
Nvidia explained that the system is designed to consume no water in cooling operations. Since the Vera Rubin architecture is capable of operating with coolants reaching up to 45 degrees Celsius, dry coolers can be relied upon instead of the traditional evaporative cooling towers prevalent in current data centers.
The company's estimates indicate that this type of cooling can save about 54 million liters of water daily for a 2-gigawatt project, although some facilities in extremely hot climates may require auxiliary chillers on the hottest days of the year.
What Changes Practically?
Unlike previous attempts at portable data centers used for transport, emergency, or military purposes by other companies such as Sun Microsystems, Amazon, and Microsoft, DSX POD containers represent essential and direct building blocks for modern AI factories.
Nvidia intends to roll out these systems globally in collaboration with 17 partners, including Armada, Dell, HPE, Schneider Electric, and Vertiv, with partners beginning to deliver the systems to the market in early 2027.