At GTC Berlin on October 21, 2026, NVIDIA announced DSX POD, a reference design in which an AI data center is assembled in a factory and delivered to the installation site in finished form. Compute equipment built on the next-generation AI platform Vera Rubin is integrated with networking, power, a liquid cooling system and more. NVIDIA says this can shorten on-site deployment, which previously took about six months, to as little as 14 days. Deliveries worldwide through partner companies are scheduled to begin in early 2027.
The large reduction comes from moving much of the work that used to be done in sequence at a construction site, including equipment installation, cabling, piping and operational testing, into the factory, where it can proceed in parallel with site preparation. However, the 14-day figure assumes installation at a site where power, cooling, networking and other infrastructure are already in place. It is not the total time including site development and manufacturing. To actually shorten the time until AI training and inference can begin, factory assembly and on-site infrastructure work have to be properly coordinated.
Factory assembly and validation cut deployment to as little as 14 days
DSX POD turns DSX, NVIDIA's design foundation for AI data centers, into a modular facility assembled in a factory. According to NVIDIA's official announcement, rack-scale compute systems based on Vera Rubin are housed in dedicated containers, and the inter-rack cabling, power wiring and cooling piping are connected inside the factory. Full-system functional testing is also performed before shipment, so that operation can begin once the unit is connected to the required external facilities at the site.
In conventional data center construction, site selection and the design of power and cooling facilities are followed by on-site installation of racks and cables, and then step-by-step validation of the network and compute systems. In this process, equipment delivery timing, the availability of specialist workers and the connection work between multiple systems affect one another, so a delay in one area easily spreads to the entire schedule. With DSX POD, much of this integration work moves into a controlled factory environment, with the aim of reducing on-site work and limiting rework caused by configuration mismatches or connection errors.
Containerized data centers have existed for some time, but DSX POD is not simply a way of putting computers in a container. In addition to Vera Rubin compute systems, networking, storage, power, cooling and control functions are designed as a single integrated facility, and the integrated state is validated before shipment. NVIDIA is aiming to shift from assembling individual racks into one system on site to delivering a pre-validated facility to the installation site.
According to NVIDIA's technical explanation, the DSX POD deployment process can be organized into three stages: factory integration and testing, site preparation carried out in parallel, and connection and validation after delivery.
| Stage / location | Main work | Relationship to shorter deployment |
|---|---|---|
| Factory, before shipment | Rack installation, internal wiring and piping, control function integration, functional testing | Moves forward integration work previously done on site |
| Installation site, in parallel with factory work | Site and foundation preparation, power, cooling, external network readiness | Proceeds alongside container manufacturing to reduce waiting time after delivery |
| Installation site, after delivery | Container placement, inter-container connections, connection to external facilities, post-transport inspection | Confirms connections and whole-system operation before moving to production use |
It is important to note that the 14-day figure does not include factory manufacturing or site preparation. NVIDIA's figure applies to installing factory-assembled equipment at a site where the necessary infrastructure is ready. It does not mean a data center can be operational 14 days after a construction plan begins. The design shortens the overall process by running factory integration and on-site preparation in parallel.
Pre-shipment testing also differs from conventional practice. In NVIDIA's reference design, the integrated equipment inside a container can be validated as a single unit rather than rack by rack. By checking equipment configuration, firmware, power-on behavior, network connections, sensors and cooling circuits before shipment, the design reduces the number of connections made for the first time on site. The aim is to lower the risk that a problem is found in the final stage of deployment and takes time to fix.
800V DC power and liquid cooling integrated, but on-site power and cooling facilities are still required
DSX POD also standardizes power and cooling. Power is based on 800V direct-current input (800V DC), and the reference design covers everything from input to the container through protection circuits, power distribution, grounding, rack connections, control and monitoring. Upstream of that, however, power generation equipment, voltage conversion equipment and the connection to the local grid are designed by partner companies to suit site conditions. Even with unified connection specifications on the container side, the unit cannot simply be plugged into power facilities anywhere in the world; local regulations and power conditions must be addressed.
Cooling is likewise divided between what is assembled in the factory and what is built on site. Because Vera Rubin rack-scale systems use liquid cooling, the container includes piping that connects racks to cooling equipment, distribution mechanisms, shutoff valves, sensors and leak detection. In the factory, in addition to assembling the liquid cooling circuit, cleaning, pressure testing and functional checks can be carried out. At the installation site, this circuit must be connected to a chilled water system and heat rejection equipment prepared in advance, and cooling performance must be confirmed under actual operating conditions.
In other words, DSX POD reduces complex integration work on site, but it does not eliminate the need for the external facilities required to run a data center. For example, even if the cooling circuit has passed pressure testing at the factory, production AI workloads cannot begin until it has been connected to the site's heat rejection equipment and operation has been confirmed. Matching the factory completion date with the date on-site power and cooling facilities become available is key to shortening deployment.
Because expensive, fully assembled compute equipment is transported inside the container, transport safety is also part of the design. NVIDIA's technical explanation says it takes into account the container structure, how equipment is secured, weight distribution, center of gravity, lifting points and handling during transport. Equipment configuration and test results are recorded before shipment, and specified inspections are repeated after arrival. The purpose is to confirm that the condition verified at the factory has not been compromised in transit; passing pre-shipment tests does not mean post-installation validation can be skipped.
Near-zero cooling water and up to 30% less power for AI processing
On the same day as the DSX POD announcement, October 21, NVIDIA also described power-saving and water-saving technologies in the broader DSX platform. According to the company, the DSX reference design can bring cooling water consumption to nearly zero, and software-based power control can achieve the same AI workload with up to 30% less power. These effects, however, stem from the overall DSX design and control technologies and should be assessed separately from the deployment-time reduction of DSX POD.
The water reduction relies on combining a liquid cooling system with dry coolers. Conventional cooling towers release heat by evaporating water, so lost water must be continually replenished. In the DSX reference design, by contrast, coolant circulates in a closed loop, and heat recovered from the compute equipment is released into the air by outdoor dry coolers. A dry cooler works much like a large radiator and can dissipate heat without evaporating water, which greatly reduces the facility's overall cooling water consumption.
According to NVIDIA, a 2GW AI data center could save 54 million liters of water per day compared with a facility using conventional cooling towers. This is an estimate for a facility's cooling water and does not mean the same reduction can be achieved everywhere. Depending on local climate, additional mechanical cooling may be needed, particularly in hot periods. Also, using coolant in a liquid cooling circuit is distinct from consuming water for heat rejection, and the liquid cooling system itself does not become unnecessary.
Power efficiency gains rely on NVIDIA's "DSX MaxLPS." This software dynamically allocates power for compute processing within the power allotted to a facility, and NVIDIA says it can cut the power needed to generate the same number of tokens by up to 30%. By allocating available power efficiently, it also aims to raise throughput without increasing power capacity. However, the up-to-30% figure is the effect NVIDIA has presented and is not a guarantee that every facility deploying DSX POD will see the same reduction.
Another component, "DSX Flex," adjusts an AI data center's power consumption according to grid conditions. By increasing or decreasing the power used for compute in response to requests from utilities, demand fluctuations and electricity prices, it helps support grid stability. According to DatacenterDynamics' October 21 report, NVIDIA's Ian Buck said at a pre-announcement briefing that when a grid needs an additional 1GW, a data center can also become a resource for meeting that demand. While DSX POD is a technology for speeding up the deployment of compute equipment, DSX MaxLPS and DSX Flex optimize power use after operations begin.
Deliveries worldwide from early 2027, with 17 partners involved
NVIDIA plans to begin delivering DSX POD worldwide in early 2027 and is working with a total of 17 partner companies. Participants include server makers such as Dell Technologies, HPE and Lenovo, as well as power and cooling equipment companies such as Schneider Electric and Vertiv. Others named are Anduril, Armada, Bull, Crusoe, DayOne, Delta, Eaton, Foxconn (Ingrasys), GIGABYTE, Pegatron, Quanta Cloud Technology (QCT) and Wistron.
NVIDIA provides the validated reference design, while partners handle actual manufacturing, regional design adjustments, customization to customer requirements and post-deployment maintenance. Server makers handle compute system integration, and specialist power and cooling companies build infrastructure suited to site conditions. Local operators also prepare the site, foundations and external facilities, and take part in container installation and pre-operation checks. Even with a common reference design, actual deployment requires coordination among multiple companies and fields of expertise.
NVIDIA's goal is to reduce the work of assembling compute equipment from scratch on site and to shorten the time until AI model training and inference can begin in a production environment. In an AI data center, even after expensive compute equipment is installed, it cannot generate services or revenue until it is operating. If delivering factory-integrated and validated equipment shortens deployment, it could allow that equipment to be put to actual computing use sooner. Conventional data center construction and on-site integration will remain options, depending on facility conditions and customer requirements.
The key question going forward is whether the 14-day deployment time NVIDIA is touting can actually be achieved in the deliveries planned for early 2027. That will require not only that factory manufacturing and validation finish on schedule, but also that on-site power, cooling, networking and other facilities are usable at the time of delivery. Systems must then pass post-arrival connection checks and whole-system validation and be able to run AI training and inference reliably. How well the work reduced by factory integration and the work remaining on site are managed will determine how effective DSX POD is at shortening deployment.