Nvidia DSX Pod: Containerized AI Data Centers Deploy in 14 Days With 504 GPUs

Nvidia announced its DSX Pod AI Factory at GTC Berlin this week, introducing a containerized solution designed to accelerate infrastructure deployment for AI data centers from six months to as little as 14 days1

1

2

2

. The modular datacenter containers represent the company's latest push to position itself as a full-stack AI infrastructure provider, moving beyond individual GPU chips to deliver complete, factory-built AI factories ready for rapid deployment.

Factory-Built Systems Slash Deployment Timelines

The Nvidia DSX Pod houses seven Vera Rubin NVL72 racks containing 504 GPUs and 252 Vera CPUs inside a shipping container roughly the size of a standard transport unit, weighing approximately 28,000 kilograms3

3

4

4

. These containerized AI data centers arrive with all computing, networking, cabling, rack-level power, and liquid cooling pre-assembled and validated at manufacturing facilities. According to Ian Buck, Nvidia's vice president of hyperscale computing, the only work required on-site involves positioning the container, connecting it to adjacent units, and completing final validation testing2

2

.

This reference design shifts the extensive integration work away from construction sites into controlled factory environments, addressing supply chain challenges and reducing exposure to component delays and costly on-site errors3

3

. The approach fundamentally changes how AI infrastructure gets built, as traditional deployments require extensive coordination across servers, networking, cooling systems, and power infrastructure.

Plug and Play Racks Target Revenue Generation

"An AI factory creates value when it starts producing tokens -- not when construction is complete," stated Jensen Huang, Nvidia's founder and CEO3

3

4

4

. The DSX Pod's design prioritizes time to first token, the critical moment when new infrastructure begins running models to serve production AI workloads and generate revenue for operators. By compressing deployment timelines from months to weeks, customers can begin earning returns on their infrastructure investments significantly faster.

The modular containers are designed for transport by truck on ordinary roads, making them accessible for deployment across diverse geographic locations2

2

4

4

. Once delivered to prepared sites, the units connect to existing power, cooling, networking, and supporting infrastructure. Nvidia emphasized that the pods are entirely customizable, allowing customers to select preferred partners for different components while maintaining the accelerated deployment timeline1

1

.

Global Manufacturing Partnership Ecosystem

Nvidia is collaborating with an extensive network of partners to manufacture and deliver DSX Pod systems globally starting in early 2027. The ecosystem includes computing partners like HPE, Dell Technologies, Lenovo, Foxconn (Ingrasys), Gigabyte, Quanta Cloud Technology, Pegatron, and Wistron, alongside infrastructure specialists such as Schneider Electric, Vertiv, Eaton, and Delta3

3

4

4

. Additional partners include Anduril, Armada, Bull, Crusoe, and DayOne.

While Nvidia has not disclosed pricing for the DSX Pod, analysts estimate that a single Vera Rubin rack can cost more than $7 million2

2

4

4

. This suggests the complete seven-rack container system represents a substantial investment, though the accelerated time to revenue generation may offset the upfront costs for operators racing to expand AI capacity.

Expanding Revenue Per Gigawatt Strategy

The DSX Pod continues Nvidia's strategic evolution from chip manufacturer to comprehensive infrastructure provider. Huang recently outlined this progression on a podcast, noting that Nvidia's offering per gigawatt of datacenter capacity has expanded from approximately $18 billion to over $40 billion2

2

. The product stack has evolved from individual GPU chips to servers with four or eight GPUs, then complete racks, and now to the DSX Pod bundling multiple racks together in transportable modules.

At GTC Berlin, Nvidia also introduced new specifications for waterless datacenters using closed-loop coolants and dry coolers rather than water-intensive cooling towers1

1

. This builds on earlier announcements focused on extracting more performance per watt while enabling facilities to adjust AI workloads based on grid conditions. Watch for how these containerized systems perform in real-world deployments and whether the 14-day timeline holds across different site conditions and infrastructure readiness levels. The modular approach could reshape competitive dynamics in AI infrastructure, particularly for operators seeking to scale capacity quickly in response to surging demand for model training and inference services.