Building an AI data center usually means months of cabling, plumbing and testing on site. NVIDIA's new pitch is to do most of that in a factory and ship the result in containers.
What NVIDIA announced
On Oct 21, 2026, at GTC Berlin, NVIDIA introduced DSX POD. It's a reference design for complete Vera Rubin AI factories, built by partners inside modular containers and validated before delivery.
The headline claim: units arrive ready to connect to a prepared site's power, cooling and networking, cutting deployment time from six months to as little as 14 days.
"An AI factory creates value when it starts producing tokens — not when construction is complete," NVIDIA CEO Jensen Huang said in the release.
What actually moves into the factory
Per NVIDIA's technical blog, partners do this before shipping:
- Install the rack-scale Vera Rubin systems inside the container.
- Rack-to-rack network cabling, power wiring, control wiring and sensors, wherever practical.
- Liquid cooling plumbing: assemble, flush, pressure-test and validate the internal loop.
- Functional tests: firmware baselines, power-on checks, fabric connectivity, cooling integrity and system-level tests.
Power comes in through a standard 800 VDC input design at the container boundary. Partners engineer the upstream power for each site, because grids and codes differ by region.
The trick is parallel work. While the containers are built and tested, the site team prepares land, foundations, power and connectivity.
Who builds it
NVIDIA lists Anduril, Armada, Bull, Crusoe, DayOne, Dell Technologies, Delta, Eaton, Foxconn (Ingrasys), GIGABYTE, HPE, Lenovo, Pegatron, Quanta Cloud Technology (QCT), Schneider Electric, Vertiv and Wistron. They're targeting delivery worldwide in early 2027.
What it doesn't fix
NVIDIA's own blog is direct about this: "DSX POD changes where integration occurs. It does not eliminate the supporting facility infrastructure."
You still need:
- A grid connection or onsite power. The IEA says grid connections, transformers and gas turbines are among the tightest bottlenecks right now. We cover that in AI data centres and the power grid.
- Water and heat-rejection infrastructure for the liquid cooling.
- The chips and memory. A container doesn't help if Vera Rubin systems or HBM are supply-limited.
The 14-day figure is NVIDIA's claim for a prepared site, not for a field. And the blog notes containerization isn't mandatory; conventional builds remain an option.
Why it matters
Faster time to first token means rented capacity comes online sooner, which feeds into GPU rental prices and into the economics of neoclouds like CoreWeave and Nebius. It also makes AI capacity look more like a product you order than a building you construct. That shift matters for how this hardware gets financed.
