NVIDIA DSX: AI factories that use less power and water, built in days

The news: DSX becomes a factory-built AI factory playbook

NVIDIA’s latest DSX announcement is not simply another server launch. The company is positioning DSX as a full-stack architecture for AI factories, and DSX POD as the modular, factory-built version of that architecture: complete Vera Rubin-based AI factory units assembled, cabled, plumbed, integrated and validated before they arrive at a prepared site [1].

The headline claim is speed. NVIDIA says DSX POD can reduce deployment at a prepared site from roughly six months to as little as 14 days, because the most repeatable and error-prone integration tasks move from the construction site into a manufacturing environment [1]. That does not mean a greenfield data center appears in two weeks. It means that once land, foundations, utility connections, heat-rejection infrastructure, networking and other site interfaces are ready, the compute container can be placed, connected, checked and commissioned much faster than in a traditional rack-by-rack build [3].

This distinction matters. NVIDIA’s own technical explanation says DSX POD changes where integration happens, not whether supporting facility infrastructure is required [3]. The product story is therefore less “instant data center” than “industrialized AI factory,” with standardized interfaces and prevalidated units designed to make capacity expansion more predictable.

A co-designed system, not a pile of components

The DSX platform is built around one main premise: compute, networking, storage, software, power and cooling should be designed as one system rather than bought and reconciled as separate layers [2]. In traditional data center projects, late-stage integration can expose mismatches between racks, switches, power distribution, cooling loops, controls and operations software. DSX tries to reduce that friction by making those layers part of a reference architecture.

DSX POD applies the same idea to physical deployment. NVIDIA says partners complete rack-to-rack cabling, wiring, piping and functional validation before shipment [1]. The technical blog adds that validation can cover rack inventory, firmware and configuration baselines, power-on checks, fabric connectivity, sensor operation, cooling integrity and system-level functional tests before the container leaves the factory [3].

That approach matters because AI infrastructure is becoming dense enough that the “server room” model no longer describes the job. A Vera Rubin DSX POD, according to NVIDIA’s GTC Berlin presentation as reported by Data Center Dynamics, starts with Vera Rubin NVL72 and combines seven liquid-cooled compute racks connected by 17 kilometers, or about 11 miles, of Spectrum-X cables [4]. At that density, cabling discipline, coolant routing, telemetry and power controls are not secondary details; they are the factory’s nervous system.

Less water: closed-loop liquid cooling and dry coolers

The environmental part of NVIDIA’s DSX message centers on cooling. Conventional cooling towers reject heat by evaporating water, which requires continual water replacement. NVIDIA says DSX reference designs connect liquid-cooled compute systems to dry-cooler heat rejection, so coolant absorbs heat from equipment and circulates in a closed loop while outdoor radiator-like systems release heat into the air without evaporating water [2].

For Vera Rubin infrastructure, NVIDIA says the design supports 100% liquid cooling with coolant temperatures up to 45 degrees Celsius. With dry-cooler-based designs, the company says facility cooling water consumption can be reduced to near zero, and at a 2-gigawatt scale it estimates savings of 54 million liters of water per day versus conventional cooling-tower systems [2].

The important caveat is climate. NVIDIA notes that local weather determines whether supplemental mechanical cooling is needed during the hottest periods [2]. In other words, DSX can sharply reduce facility cooling water use, but deployment details will depend on geography, design temperature, operating envelope and the economics of dry cooling versus additional mechanical systems.

Less power per token: DSX MaxLPS and flexible loads

NVIDIA is also making a power-efficiency argument. Its DSX MaxLPS software dynamically allocates power inside an AI factory’s provisioned budget, with the goal of producing the same number of tokens using up to 30% less power [2]. In a Lambda proof of concept across NVIDIA HGX B200 systems, NVIDIA says DSX MaxLPS delivered 24% more cluster token throughput and 19% more GPU capacity within a fixed power budget [2].

The company also highlighted an Nscale evaluation in Iceland across four NVIDIA GB300 NVL72 racks. In that test, NVIDIA says DSX MaxLPS produced 49% higher inference throughput and 37% more managed GPU capacity within the same 264-kilowatt provisioned power budget [2]. These figures are vendor-reported and workload-specific, but they show the metric NVIDIA wants the market to focus on: not megawatts alone, but tokens produced per constrained megawatt.

DSX Flex extends the argument from facility efficiency to grid behavior. NVIDIA says DSX Flex helps AI factories adjust workload power consumption in response to utility needs, demand-response events and electricity prices, while coordinating grid power, onsite renewables and storage [2]. In an early trial with Emerald AI, EPRI, Nebius, the U.K. National Grid and NVIDIA, the companies demonstrated power-draw reductions of up to 40% on an NVIDIA Blackwell Ultra GPU cluster during simulated grid events while critical workloads continued running [2].

Why Europe is central to the announcement

NVIDIA framed these announcements at GTC Berlin, and the European context is not incidental. AP reported that Jensen Huang used the Berlin event to urge Europe’s AI sector to move faster and described modular AI data centers as a way to address the region’s compute shortage [5]. AP also reported that Europe has about 5% of global AI computing capacity, compared with about 75% in the United States, making compute availability a strategic bottleneck for European developers [5].

That helps explain why NVIDIA’s DSX message combines speed, efficiency and grid coordination. Europe’s AI challenge is not only obtaining GPUs; it is connecting power, cooling, construction, permitting, industrial partners and operations software into productive capacity. DSX is NVIDIA’s attempt to package that whole chain into an architecture partners can replicate.

NVIDIA says it is working with Anduril, Armada, Bull, Crusoe, DayOne, Dell Technologies, Delta, Eaton, Foxconn, GIGABYTE, HPE, Lenovo, Pegatron, QCT, Schneider Electric, Vertiv and Wistron to deliver DSX POD AI factories worldwide in early 2027 [1]. The partner list shows that DSX POD is meant to be an ecosystem model, not a single NVIDIA-built box.

Digital twins and operations complete the stack

DSX is also a software and simulation story. NVIDIA says DSX Sim and the Omniverse DSX Blueprint give developers reference digital-twin assets, scripts and workflows for building AI factory models on OpenUSD and Omniverse libraries [6]. The goal is to evaluate power, cooling, equipment placement and capacity choices before hardware is installed.

That is consistent with the broader DSX concept: the AI factory is treated as a system throughout design, construction and operations. NVIDIA says it is expanding DSX OS with components such as Config Manager, Rack Management Service and Cluster Readiness Engine to automate network configuration, manage rack-level inventory and power, and validate clusters before production workloads arrive [2].

Armada’s announcement shows how partners may translate the model into commercial offerings. The company said it will support the NVIDIA DSX POD reference design with Johnson Controls, engineering and manufacturing prefabricated superintelligence factories based on Vera Rubin, and it plans 500 megawatts of distributed sovereign capacity in 2027 [7].

The bottom line

DSX POD is best understood as NVIDIA’s bid to industrialize AI infrastructure. It does not remove the need for land, power, cooling infrastructure, permitting or careful engineering. But it moves a major part of the integration burden into the factory, standardizes the interfaces, and treats performance, power, water and operations as one co-designed system.

If the claims hold at scale, the payoff is significant: faster time to first production tokens, much lower facility cooling water use in suitable climates, and more AI output from constrained power budgets. The open question is execution. DSX POD’s promise will depend on partner manufacturing quality, local grid conditions, thermal design, workload profiles and the ability to deliver these systems globally in early 2027 without turning “modular” into another supply-chain bottleneck [1].

Sources from the last 72 hours

1. [1]NVIDIA Introduces DSX POD AI Factories to Accelerate Europe’s Infrastructure DeploymentOct 21, 2026, 2:00 AM
2. [2]European Ecosystem Builds World’s Most Efficient AI Factories With NVIDIA DSXOct 21, 2026, 2:00 AM
3. [3]Achieve Productive AI Capacity Faster and More Predictably with NVIDIA DSX PODOct 21, 2026, 2:00 AM
4. [4]Nvidia launches DSX POD, claims the Vera-Rubin reference architecture will make AI data centers more water and energy efficientOct 21, 2026, 2:00 AM
5. [5]Developing an AI Factory Digital Twin with an Agent-Assisted WorkflowOct 21, 2026, 2:00 AM
6. [6]Armada Unveils Superintelligence Factories Built Within NVIDIA DSX POD, Deployable in Months, Not YearsOct 21, 2026, 1:30 PM
7. [7]Nvidia CEO cheerleads for Europe's nascent AI sectorOct 21, 2026, 2:28 PM

AI-generated article based on recent web research, then preserved as a dated editorial snapshot.