Shipping AI Factories in Containers
Nvidia unveiled the DSX POD at GTC Berlin. Vera Rubin AI factories are assembled and verified in factories, then shipped in containers. It is estimated that deployment on a prepared site will be reduced from six months to 14 days.
AI Summary
Nvidia unveiled the DSX POD at GTC Berlin, a reference design that packages fully assembled Vera Rubin AI factories into modular containers. By transforming data center construction into a streamlined manufacturing process, Nvidia estimates deployment times can be reduced from six months to just 14 days on prepared sites. This shift fundamentally relocates the industry's bottleneck, making access to sites with pre-installed power, cooling, and networks the ultimate competitive advantage.
Nvidia unveiled the DSX POD during its keynote speech at GTC Berlin on October 21. Instead of building data centers on-site, the approach involves manufacturing them in factories and shipping them in containers.
The configuration is a reference design that packages an entire Vera Rubin-based AI factory into a modular container. Rack-to-rack wiring, power connections, cooling piping, and functional testing are all completed at partner factories before being sent to the site. The only tasks left at the site are connecting the container to power, cooling, and networks.
The number Nvidia highlighted is 14 days. The company stated that if a site is equipped with power and cooling, the deployment period is reduced from six months to a minimum of 14 days. CEO Jensen Huang explained that an AI factory begins delivering value not when construction finishes, but when it starts generating tokens, noting that this brings factory-made precision and speed to AI infrastructure construction.
| Conventional On-site Deployment | DSX POD | |
|---|---|---|
| Cabling & Plumbing | On-site work | Completed at partner factory |
| Functional Verification | On-site commissioning | Pre-shipment verification |
| Deployment Time | Approx. 6 months | Minimum 14 days |
| On-site Tasks | Virtually everything | Power, cooling, and network connections |
The supplier list features 17 companies. Server and EMS providers like Dell Technologies, HPE, Lenovo, Foxconn (Ingrasys), Quanta (QCT), Pegatron, Wistron, and GIGABYTE are grouped with power and cooling specialists such as Schneider Electric, Vertiv, and Eaton. Joining them are Crusoe, Day One, Armada, Anduril, Bull, and Delta. This setup bundles server manufacturers with electrical and HVAC companies into a single product.
What is absent from the announcement is also clear. The number of racks, the number of GPUs, power capacity, and price were not disclosed. Some foreign media reported a configuration of seven Vera Rubin NVL72 racks, but this is not stated in Nvidia's official materials. The supply timeline was only given as a target of early 2027, with a caveat attached that additional cooling may be necessary depending on the climate.
The container arrives in 14 days. However, building the site to plug that container into still takes several years.
What this announcement changes is the location of the bottleneck. The difficulties surrounding AI data centers had already shifted from equipment procurement to construction, permitting, and securing power. Nvidia has replaced construction with a manufacturing problem. This narrows the remaining competitive edge down to a single factor: who holds the sites with pre-installed power, cooling, and network lines.
This article was automatically translated from the Korean original by AI. For the authoritative version, read it in Korean.
한국어 원문 읽기 →