NVIDIA DSX POD Puts 504 Rubin GPUs in a Container, Live in 14 Days

Each container weighs about 28 metric tons, The Register reports. Rack power, rack-to-rack cabling and liquid cooling are installed at the factory rather than on site, which NVIDIA says cuts exposure to component delays and error-prone field work.

Buyers can pick their vendor stack, for example HPE rather than Dell, or Eaton rather than Schneider for electrical gear, and add capacity by placing more containers. At the same event NVIDIA set out DSX specifications for waterless data centers that use closed-loop coolant and dry coolers instead of evaporative cooling towers.

• Contents: 7 Vera Rubin NVL72 racks, 504 GPUs, 252 Vera CPUs
• Weight: about 28 metric tons per container
• Deployment: as little as 14 days at a prepared site, against about six months onsite (NVIDIA claim)
• Done before shipment: cabling, wiring, piping, rack power, liquid cooling and functional validation
• Availability: partner deliveries from early 2027

Our arithmetic: if a typical onsite build takes about 180 days, the claimed 14 days cuts deployment time by more than 90%. The catch is the phrase prepared site. Grid connection, substation work and cooling plant are still the customer's job, and those are the slowest parts of most projects we cover.

Voltgrid Analysis: Infrastructure Implications

Our read: NVIDIA is trying to fix a problem it does not control. Chips are no longer the slowest part of an AI build; sites are. Microsoft's decision to drop an Nscale site when UK grid power slipped to the 2030s shows how a finished hardware order can still wait years for a connection.

Containers will not shorten that queue, but they shrink everything after it. A site with power available could start producing within weeks, and a buyer could add capacity one container at a time instead of committing to a whole hall. Our inference: that suits neoclouds and sovereign AI projects in Europe, the market NVIDIA chose for the launch.

Power density is the open question. The material we reviewed gives no per-container power figure, and seven NVL72 racks in one box will need a large, liquid-cooled electrical feed. Partners are working on that side too, as with Schneider's software-defined switchgear.

Voltgrid Verdict

Looking Ahead

Partners plan to start delivering DSX POD units in early 2027. The first deployments will show whether the 14-day figure holds outside NVIDIA's reference conditions.

Source: NVIDIA, The Register.