DSX POD aims to reduce AI factory deployment to 14 days

Nvidia presented DSX POD at GTC Berlin, a container-mounted AI data center design factory-prepared. The promise: reduce deployment from six months to 14 days on a prepared site, enabling faster production start.

On October 21, 2026, at GTC Berlin, Nvidia unveiled DSX POD, a reference design for Vera Rubin "AI factories" assembled in modular containers and validated before delivery. The goal is simple: to reduce on-site installation time from six months to 14 days on a prepared site, in order to launch AI computing workloads into production faster and generate revenue sooner.

An AI factory arriving almost ready to plug in

In the model described by Nvidia, the containerized units arrive ready to be connected to power, cooling, and the network of an already prepared site. The integration of computing, wiring, rack-level power, and liquid cooling is done in the factory, then partners deliver a tested set rather than a series of components to be assembled under pressure on site.

This approach aims to reduce construction delays, the need for specialized labor, and logistical uncertainties. Nvidia also wants to limit delays related to components and on-site operations, which are often more costly and prone to errors. In other words, the idea is to shift part of the industrial risk off-site, where delays tend to drag on.

Why Nvidia is pushing this format now

Jensen Huang, founder and CEO of Nvidia, sums up the logic in a phrase: "An AI factory creates value when it starts producing tokens — not when construction is complete," he said. Nvidia defines "time to first token" as the moment when a new AI factory begins running models for production workloads. In this view, every day gained in deployment directly impacts the site's economics.

The message is aimed at operators deploying heavy infrastructure for training, then inference, that is, running already trained models. Nvidia highlights a faster transition from construction to monetization, with partners able to manufacture, validate, and ship these blocks at scale.

DSX, the software and power brick behind the hardware

DSX POD is part of a broader platform, DSX, which Nvidia describes as a full-stack reference architecture co-designed with infrastructure players. The system combines computing, networking, storage, software, power, and cooling as a single unit, with components like DSX MaxLPS, DSX Flex, DSX Sim, DSX OS, and DSX Exchange.

On the power side, Nvidia claims that DSX MaxLPS can dynamically allocate power within the available budget, with up to 30% less power used to produce the same number of tokens. The company also cites a proof of concept conducted by Lambda on HGX B200 systems, with 24% more token throughput and 19% additional GPU capacity within a fixed power budget, according to the announcement. These figures remain tied to the test framework provided by Nvidia.

Cooling becomes a selling point

The platform also highlights liquid cooling systems combined with heat dissipation via dry coolers, which can reduce cooling water consumption to nearly zero. Nvidia even states that its systems can use up to 100% less cooling water, although some climates will still require supplementation during periods of high heat [to be verified].

This aspect matters as much for operating costs as for environmental and regulatory constraints. In several markets, electrical availability, water management, and the ability to cool very high computing densities are becoming deployment criteria as important as GPU performance itself.

An already aligned partner ecosystem

Nvidia names Anduril, Armada, Bull, Crusoe, DayOne, Dell Technologies, Delta, Eaton, Foxconn (Ingrasys), GIGABYTE, HPE, Lenovo, Pegatron, Quanta Cloud Technology (QCT), Schneider Electric, Vertiv, and Wistron among the partners expected to deliver these DSX POD AI factories. The first global deliveries are announced for early 2027.

This strategy also reflects a market evolution: large AI centers are no longer sold just as servers, but as industrialized systems, integrated upstream and delivered almost as finished products. For customers, the benefit is clear; for infrastructure providers, the competition shifts to the ability to assemble quickly, validate flawlessly, and keep the commissioning promise.

Key points

  • DSX POD was presented on October 21, 2026, at GTC Berlin.
  • Deployment time is reduced from six months to 14 days on a prepared site.
  • Units are preassembled and tested in the factory before delivery.
  • Nvidia announces global deliveries starting early 2027.
  • Lambda served as proof of concept on HGX B200.
  • Jensen Huang speaks of token production before construction completion.

By the numbers

  • 14 days — announced deployment time on a prepared site, Nvidia announcement, October 21, 2026.
  • 6 months — commissioning time mentioned for traditional deployment, Nvidia announcement, October 21, 2026.
  • 30% — power reduction announced with DSX MaxLPS, according to Nvidia.
  • 24% — token throughput increase cited in Lambda proof of concept, according to Nvidia.
  • 19% — additional GPU capacity announced in the same test, according to Nvidia.

To read