Nvidia introduced DSX POD on 21 October, saying its factory-built Vera Rubin AI factories could be deployed at a prepared site in as little as 14 days, against six months for an on-site build. The company announced the modular reference design in a release at GTC Berlin. Partners would assemble and test the systems in containers before delivery, shifting integration work away from the customer’s site.
Key points
- Nvidia says its prefabricated containers could shorten installation to as little as 14 days at a site ready to receive them, versus six months for a conventional build.
- Its DSX cooling design could save 54 million liters of water a day at 2-gigawatt scale, compared with conventional cooling towers, the company says.
- Lambda demonstrated 24% more cluster token throughput in a DSX MaxLPS proof of concept, according to Nvidia.
- Nvidia is working with companies to deliver DSX POD factories worldwide in early 2027.
DSX POD shifts installation to the factory
DSX is Nvidia’s reference architecture for AI factories, bringing compute, networking, storage, software, power and cooling into one design. The platform was first unveiled at GTC San Jose earlier in 2026, Data Center Dynamics reported. DSX POD applies a container-based design to Vera Rubin infrastructure.
Under the POD design, partners would complete cabling between racks, wiring, piping and functional checks before shipment, Nvidia says. The containers would then connect to power, cooling, networking and other infrastructure at a prepared site. Nvidia says site preparation and factory integration can take place separately, reducing the on-site construction and installation work required before the systems run production workloads.
The 14-day deployment figure applies to a site that is already prepared. Nvidia ties the shorter installation period to an earlier start for production inference and model training. Jensen Huang, Nvidia’s founder and chief executive, said the aim was to bring capacity into revenue-generating use sooner.
The design also gives Nvidia’s computing and infrastructure partners a repeatable configuration to build and deliver. It extends the company’s DSX approach from specifying the systems that go into an AI factory to specifying how they are integrated and checked before they reach the operator.
DSX cooling targets water use at 2 gigawatts
In a separate 21 October release, Nvidia said DSX reference designs combine closed-loop liquid cooling with outdoor dry coolers. Liquid carries heat away from computing equipment, while the dry coolers release it into the air rather than evaporating water as conventional cooling towers do.
Nvidia says Vera Rubin infrastructure supports full liquid cooling at coolant temperatures of up to 45 degrees Celsius. With dry coolers, facility water consumption for cooling can fall to near zero, according to the company. It estimates that a DSX installation at 2-gigawatt scale could save 54 million liters of water a day against conventional cooling-tower systems, an amount it equates to the daily use of 180,000 households. Local climate can require supplemental mechanical cooling during the hottest periods, Nvidia says.
Cooling equipment is also part of the wider supplier effort around DSX. AI Affairs reported that LG Electronics qualified a 2.6-MW cooling unit for Nvidia AI factories.
DSX MaxLPS works within power budgets
Nvidia says DSX MaxLPS allocates power dynamically within an AI factory’s provisioned budget. The company says operators could produce the same number of tokens with up to 30% less power, or use the available budget to support more GPU capacity and throughput. CoreWeave, Firmus, Google Cloud, IREN, Lambda, Nebius, Nscale and Sharon AI are exploring the software, according to Nvidia.
Nvidia says Lambda recorded gains of 24% in cluster token throughput and 19% in GPU capacity under the same power limit in a DSX MaxLPS proof of concept using Nvidia HGX B200 systems. DSX Flex, another part of the platform, is designed to adjust workload power consumption in response to utility needs, demand-response events and electricity prices.
Nvidia says an early DSX Flex trial with Emerald AI, EPRI, Nebius and the UK National Grid reduced a Blackwell Ultra GPU cluster’s power draw by up to 40% in response to simulated grid events.
Nvidia is working with Anduril, Armada, Bull, Crusoe, DayOne, Dell Technologies, Delta, Eaton, Foxconn (Ingrasys), GIGABYTE, HPE, Lenovo, Pegatron, Quanta Cloud Technology, Schneider Electric, Vertiv and Wistron to deliver DSX POD AI factories worldwide in early 2027.
