Building an AI data center can take months, in part because its computing equipment, networking, power systems, cabling, and cooling must come together at the site. Nvidia’s DSX POD proposal shifts that work into factories: partners assemble and test the systems inside containers, which are then connected at a prepared location. Nvidia says the approach could cut setup time from six months to as little as 14 days.
Move assembly and testing off site
Unveiled at GTC Berlin, DSX POD is a reference design for modular data centers built around Nvidia’s Vera Rubin systems. It sets out a way for partners to prepare the equipment as a packaged unit rather than assemble the full system at its final location.
Partners such as Dell, HPE, Lenovo, and Schneider Electric install the computers, networking gear, cabling, power supply, and liquid cooling in containers at the factory. They also test the units there. The proposal therefore brings several parts of data center preparation together before the equipment reaches its destination.
At the site, the containers need to be connected to power, cooling, and networking that are already prepared. That leaves important work for the location itself, but moves the installation and testing described by Nvidia into the factory phase.
What a shorter setup could mean
Nvidia says the process could reduce setup time from six months to as little as 14 days. The difference matters to operators because the facility must be ready to run AI models before it can begin generating revenue. Nvidia’s argument is that a shorter path to operation could bring that point forward.
The proposal links the timeline to a change in where work happens. Factory assembly and testing allow the equipment to arrive as a prepared unit, while the site must provide the connections it needs. The shorter estimate depends on that prepared location and the work completed before delivery; the announcement does not say that every project will take 14 days.
For operators considering the design, the factory work and site readiness are both part of deployment. The containers carry the assembled systems, but they still need access to power, cooling, and networking at their destination. The approach changes the sequence and location of preparation rather than removing the need for infrastructure.
Cooling claims come with a climate caveat
Nvidia also says the broader DSX platform can use up to 100 percent less cooling water. Its closed-loop cooling systems release heat into the air without evaporation, which is the basis for the company’s water-use claim.
The company acknowledges that some climates will still need extra cooling during extreme heat. That caveat matters because the platform’s cooling approach does not mean every location can avoid additional cooling in all conditions. Operators would need to account for the climate at a prepared site when planning how the system will run.
Worldwide shipments are planned for early 2027
The first DSX POD units are set to ship worldwide in early 2027. That gives partners and prospective operators a stated timeframe to plan around, while the reported setup and water-use figures remain Nvidia’s claims about the design.
Together, the modular format, factory testing, and closed-loop cooling describe Nvidia’s effort to package more of the data center build before it reaches the site. The central promise is faster deployment: bring in the prepared equipment, connect it to site infrastructure, and begin running AI models sooner. Whether a specific project reaches the shortest timeline will depend, at minimum, on having its power, cooling, and network connections ready.