Nvidia unveils DSX POD server racks to speed up AI factory buildouts

Nvidia wants data centers to stop being slow construction projects and start behaving like appliances you plug in. At its GTC Berlin event, CEO Jensen Huang introduced the DSX POD, a reference design that packs seven liquid-cooled server racks into a single container-sized unit.

The pitch is speed. Nvidia claims prepared sites can hook these pods up to existing power and cooling in as little as 14 days, compared with a process that previously took up to six months.

What’s inside the box

Each DSX POD holds seven Vera Rubin NVL72 racks. Together, that adds up to 504 Rubin GPUs and 252 Vera CPUs.

The whole thing fits into an enclosure roughly the size of a standard shipping container. It also weighs around 28,000 kg.

Inside, Nvidia has threaded 17 km of Spectrum-X networking cable and built in liquid cooling. Spectrum-X is the company’s Ethernet networking platform, the plumbing that lets thousands of chips talk to each other fast enough to act like one giant computer.

Nvidia says deliveries are expected to begin in early 2027. The company is working with 17 ecosystem partners on the rollout, including Dell Technologies, HPE and Lenovo.

The math Huang wants buyers to do

Huang’s keynote leaned heavily on the idea that data centers are turning into “AI factories.”

He put numbers on it. According to Huang, a 1 GW AI factory costs approximately $50 billion to build and can generate around $40 billion in annual revenue.

Then came the stick to go with the carrot. Huang warned that a three-month deployment delay could translate into roughly $10 billion in lost revenue opportunity.

Nvidia is also selling efficiency. The company says the design can deliver up to 30% less power usage through dynamic power allocation. It also says closed-loop liquid cooling can nearly eliminate water consumption for cooling.

Closed-loop cooling works a bit like a car radiator. The coolant circulates through the system and gets reused, rather than evaporating off into the sky the way water does in many traditional cooling towers.

Why Europe got the spotlight

The venue was not an accident. Huang spent part of his keynote stressing how important AI technology is for European countries and companies.

Europe has been pushing for AI sovereignty, meaning governments want computing infrastructure built and controlled locally rather than rented from overseas providers.

What this means for the AI infrastructure race

There are real caveats. The 14-day timeline applies to prepared sites, which still need adequate power and cooling infrastructure in place.

The revenue figures also deserve some skepticism. Huang’s estimate of around $40 billion a year from a 1 GW facility depends on sustained demand for AI computing.

The timing is worth watching too. Shipments are expected to start in early 2027, so the real test of Nvidia’s speed claims will come when the first customers actually try to switch these pods on.