# NVIDIA DSX POD

> Source: https://aiwiki.ai/wiki/nvidia_dsx_pod
> Updated: 2026-10-25
> Fact-checked: 2026-10-25
> Categories: AI Hardware, AI Infrastructure, Data Centers, NVIDIA
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "NVIDIA DSX POD." aiwiki.ai, 25 Oct 2026. https://aiwiki.ai/wiki/nvidia_dsx_pod
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**NVIDIA DSX POD** is a reference design from [NVIDIA](https://aiwiki.ai/wiki/nvidia) for building complete AI data centers, which NVIDIA calls AI factories, out of factory-built shipping-container units. Partner manufacturers install NVIDIA [Vera Rubin](https://aiwiki.ai/wiki/nvidia_vera_rubin) rack-scale systems inside the containers, cable, plumb and test them at the factory, and ship the finished units to a site that already has power, cooling water and network connections in place.[1][2] NVIDIA introduced the design at its GTC Berlin conference on October 21, 2026, and said a global group of 17 computing and infrastructure partners would deliver DSX POD AI factories worldwide in early 2027.[1] The company's headline claim is that the approach cuts deployment time at a prepared site from six months to as little as 14 days; that figure is NVIDIA's own estimate and covers the work after site preparation, not land, permitting or grid connection.[1][3][13] NVIDIA itself does not build or sell the containers. It supplies the design and its validation, and partners manufacture, customize, deliver and support the systems.[2][3][9]

DSX POD is one implementation of the broader [NVIDIA DSX](https://aiwiki.ai/wiki/nvidia_dsx) platform, NVIDIA's framework for co-designing the compute, networking, storage, power, cooling, facility controls and software of a large data center.[2]

| Attribute | Detail |
|---|---|
| Developer | NVIDIA (reference design); partners build the units[1][3] |
| Type | Factory-integrated, containerized AI data center reference design[2][3] |
| Part of | [NVIDIA DSX](https://aiwiki.ai/wiki/nvidia_dsx) AI Factory Platform[2] |
| Announced | October 21, 2026, GTC Berlin keynote[1][4] |
| Compute platform | NVIDIA Vera Rubin rack-scale systems; seven Vera Rubin NVL72 racks per POD, per NVIDIA's keynote[2][14] |
| Power input | Standard 800 VDC input to the core containers[2] |
| Cooling | Liquid cooling inside the container, connected on site to prepared water and heat-rejection infrastructure[2] |
| Reported weight | About 28,000 kg (nearly 62,000 lb), per NVIDIA as reported by CNBC[6] |
| NVIDIA deployment claim | Six months to as little as 14 days at a prepared site[1] |
| Initial target customers | Large-scale, greenfield, liquid-cooled AI infrastructure, including NVIDIA Cloud Partners[3] |
| Partners named at launch | 17 (9 ODMs and 8 container solution providers on NVIDIA's product page)[1][3] |
| Planned availability | Partner deliveries worldwide in early 2027 (as of October 25, 2026)[1] |

## Background

### Prefabricated and containerized data centers

Building data center capacity from factory-made modules is not new. In 2006 Sun Microsystems presented Project Blackbox, a data center packed into a 20-foot shipping container that the company said could be deployed almost anywhere with water and electricity to plug into. The Register reported at the time that Sun expected the product to be commercially available in the second half of 2007, starting at $500,000, and that Sun claimed pre-configuration could cut installation times by as much as 90 percent.[18]

The industry's terms for such systems overlap. A Schneider Electric white paper by Wendy Torell, "Types of Prefabricated Modular Data Centers," defined a prefabricated modular data center as one made up of at least one pre-engineered, factory-integrated and pre-tested assembly of subsystems that have traditionally been installed separately on site. The paper also separated pre-engineered systems (designed to meet predetermined performance specifications and documented with, at minimum, a materials list, system-level specifications and drawings) from prefabricated ones (preassembled, integrated and tested in a factory), and gave a reference design as an example of something pre-engineered but not necessarily prefabricated.[19] By that terminology, a reference design such as DSX POD is pre-engineered, while the partner-built containers made from it are prefabricated units.

NVIDIA's blog post announcing the design acknowledges that "modular and prefabricated data centers already exist." Its argument is that an AI factory is harder to containerize than conventional IT equipment, because rack-scale AI systems combine high-density compute, scale-up and scale-out networking, liquid cooling, power delivery, telemetry and management software that behave as one system, and whose mechanical, electrical, thermal and network requirements must stay within specification through factory integration, transport, installation and operation.[2]

### Moving integration into the supply chain

NVIDIA presents DSX POD as the next step in moving assembly away from the customer site. The technical blog describes a progression from chips and boards to complete rack-scale systems and large reference architectures such as [NVIDIA DGX SuperPOD](https://aiwiki.ai/wiki/nvidia_dgx_superpod).[2] In his keynote, [Jensen Huang](https://aiwiki.ai/wiki/jensen_huang) said that with the NVLink 72 rack-scale computer NVIDIA had taken a rack that would otherwise have been assembled on site and built it inside the supply chain, and he called DSX POD "NVLink 72 now into a large cabinet system."[14] Futurum Group described the product as extending the logic of NVL72 to the scale of a multi-rack pod.[11] ServeTheHome noted that DSX POD should not be confused with DGX SuperPOD, which is a blueprint for large DGX system clusters rather than a shipped, pre-assembled container.[9]

## Announcement

NVIDIA announced DSX POD in a press release dated October 21, 2026, titled "NVIDIA Introduces DSX POD AI Factories to Accelerate Europe's Infrastructure Deployment," and in Huang's [GTC](https://aiwiki.ai/wiki/gtc) Berlin keynote the same day.[1][4] The release describes "a reference design for complete NVIDIA Vera Rubin AI factories built by partners inside modular containers, integrated and validated before delivery," and quotes Huang: "An AI factory creates value when it starts producing tokens, not when construction is complete."[1] NVIDIA published a technical blog post by Premal Savla, a senior director of product management who leads DGX Systems and Solutions, on the same date, and the NVIDIA AI Infrastructure account promoted it on X on October 22.[2][5]

A keynote video described the configuration: "It starts with Vera Rubin NVL72. Seven liquid cooled compute racks connected with 17 kilometers of Spectrum-X cables form a Vera Rubin DSX POD," with dry coolers replacing evaporative cooling and 800 volts DC connecting to the racks.[14][10] Each Vera Rubin NVL72 rack combines 72 Rubin GPUs and 36 Vera CPUs, so seven racks hold 504 GPUs, the figure Futurum Group cited.[17][11] CNBC reported that NVIDIA described the unit as roughly the size of a standard shipping container, weighing about 28,000 kilograms and designed to travel by truck on normal roads.[6] ServeTheHome wrote that NVIDIA had not confirmed the container size and that it suspected a 40-foot unit, and wrote that multiple PODs would fill out a data center "25 EFLOPs of AI inference at a time."[9] These size and performance figures do not appear in NVIDIA's press release or technical blog.

In a media briefing, Ian Buck, NVIDIA's vice president of hyperscale and high-performance computing, called it "a fully assembled, fully tested, fully pre-plumbed, prebuilt, pre-validated container that is the building block for AI factories." Buck said the design was meant for worldwide deployment and for organizations bringing compute online at the scale of hundreds of megawatts, and that NVIDIA chose Berlin because it sees modular, faster-to-deploy infrastructure playing an important role in Europe's expansion of AI capacity.[7] CNBC quoted Buck saying that on arrival a POD only needs to be "placed in its location, connected to the rest of the other containers, and validated through final burning and testing."[6] NVIDIA did not announce a price.[6]

## Design

NVIDIA's technical blog describes DSX POD as "a standardized AI factory architecture composed of factory-built container types and defined interfaces between them," designed around the Vera Rubin platform and its rack-scale systems.[2] The reference design covers a complete AI factory rather than a compute enclosure alone: compute, service and management functions, scale-out networking, storage, power, cooling, cabling and supporting infrastructure.[2] The blog groups the requirements into six areas.

| Area | What the reference design defines (per NVIDIA)[2] |
|---|---|
| Compute integration | A compute container for Vera Rubin rack-scale infrastructure; how compute and supporting racks fit in the enclosure; structural support, weight distribution, equipment restraint, service clearances, airflow and liquid-cooling access, rack-level power and network connectivity |
| Networking and storage | Connectivity inside a compute container, between containers and to the customer's external network, including scale-out fabrics, service and management networks, fiber entry points, trunk cabling, routing, labeling, bend radius and pre-shipment link validation; storage capacity, performance, connectivity and management requirements, with equipment left to the partner solution |
| Cabling and controls | Rack-to-rack network cabling, power wiring, control wiring and sensor integration moved into the factory where practical, with defined cable paths, separation, terminations, labeling, test procedures and serviceability |
| Power delivery | A standard 800 VDC input to the core containers, continuing through protection, distribution, grounding, rack connections, controls and telemetry; partners engineer upstream generation, conversion and the regional electrical solution |
| Liquid cooling | In-container plumbing from racks to the container cooling interface, including routing, manifolds, isolation, instrumentation, leak detection, pressure integrity and service access |
| Mechanical design and transport | Container structure, restraints, weight distribution, center of gravity, lifting and handling points, access paths and transport conditions, tied to factory quality gates |

### Compute and validation

Factory integration changes what gets tested before shipment. Instead of shipping each rack separately for assembly into a cluster on site, partners validate the installed compute environment as one deployment unit. NVIDIA lists rack inventory, firmware and configuration baselines, power-on checks, fabric connectivity, sensor operation, cooling integrity and system-level functional tests as items the validation plan can cover before a container leaves the factory.[2]

### Power

The power design separates what NVIDIA standardizes from what partners adapt locally. NVIDIA defines the power characteristics and interfaces the installed AI infrastructure needs, starting at a standard 800 VDC input, while partners engineer upstream generation, conversion and the regional electrical solution. The blog's reasoning is that utility conditions, electrical codes, equipment availability and grid architectures differ by region, so a common interface at the container boundary allows repeatable core designs.[2] NVIDIA's product page also points to [NVIDIA DSX MaxLPS](https://aiwiki.ai/wiki/nvidia_dsx_maxlps), its power-management technology, for policy-driven power control, workload orchestration and telemetry within fixed power limits.[3]

### Cooling

Vera Rubin rack-scale systems need coordinated liquid-cooling infrastructure. Partners can assemble, flush, pressure-test and validate the container's internal cooling loop at the factory; on site they connect it to prepared water and heat-rejection infrastructure, commission the full loop and verify performance under operating conditions. The blog states that DSX POD "changes where integration occurs. It does not eliminate the supporting facility infrastructure."[2] The keynote video paired the POD with outdoor dry coolers in place of evaporative cooling, an element that The Next Platform noted is part of the standard DSX architecture rather than new to DSX POD.[14][10] Separately, NVIDIA said the DSX AI Factory Platform as a whole can cut cooling water consumption by up to 100 percent and deliver the same AI output with up to 30 percent less power; those are platform-level company claims, not DSX POD measurements.[4][8]

### Deployment sequence

The factory-built model is meant to run site work and equipment integration in parallel. Site teams prepare land, foundations, power and external connectivity while partners install the racks, complete internal cabling and plumbing, integrate controls, load approved configuration baselines and run functional tests at the factory. Partners inspect and document each unit before shipment, secure it for transport and repeat defined checks on arrival. On site, the remaining work is placement, inter-container connections, hookup to external power, cooling and networking, and final commissioning.[2] NVIDIA's blog states that the model "does not remove onsite work," and its product FAQ says customers still need a prepared site with connections to power, facility water, cooling and network infrastructure.[2][3]

NVIDIA also argues that a standardized container gives operators a repeatable unit for expanding capacity and a known hardware environment on which software and operational practices can be tuned, while partners adapt each implementation to local codes, site conditions and supply chains.[2] The design does not make containerization mandatory; NVIDIA describes conventional construction and onsite-integrated designs as remaining options, with DSX POD aimed at large deployments where schedule predictability and reduced onsite work are priorities.[2][3]

## Deployment-time claims

NVIDIA's figures for time savings differ by source, and all of them are company estimates rather than results from a completed, independently measured deployment.

| Source | Claim |
|---|---|
| Press release, October 21, 2026 | Deployment time at a prepared site falls "from six months to as little as 14 days"[1] |
| Huang's quote in the press release | Deployment time reduced "from months to weeks"[1] |
| Keynote video | "Once site utilities are ready, estimated on-site deployment goes from several months to days"[14] |
| Huang, speaking in the keynote | Time from installation and arrival of computing to production of tokens goes from "something along the lines of four to five months down to about a month"[14][10] |

Huang tied the timing to cost. In the keynote he said a 1-gigawatt data center costs about $50 billion and can generate about $40 billion in revenue, so a three-month delay costs about $10 billion in revenue.[14][10]

Several analysts pointed out what the 14-day figure leaves out. Constellation Research noted that the catch is the need for a prepared site, which also needs power.[12] Converge Digest wrote that the figure excludes potentially lengthy land, permitting, construction and utility-power work, and that the announcement did not establish independently validated installation timelines, detailed power density or total deployment cost.[13] Futurum Group argued that grid-connection restrictions and multi-year queues in Europe's main data center markets dwarf the six months that DSX POD removes.[11]

## Partners

NVIDIA's press release lists 17 companies working with it to deliver DSX POD AI factories: Anduril, Armada, Bull, Crusoe, DayOne, Dell Technologies, Delta, Eaton, Foxconn (Ingrasys), GIGABYTE, HPE, Lenovo, Pegatron, Quanta Cloud Technology (QCT), Schneider Electric, Vertiv and Wistron. The release subtitle highlights Bull, Dell Technologies, HPE, Lenovo, Schneider Electric and Vertiv.[1] NVIDIA's product page sorts them into two groups.[3]

| Group (NVIDIA product page)[3] | Companies |
|---|---|
| ODMs | Bull, Dell, GIGABYTE, HPE, Ingrasys, [Lenovo](https://aiwiki.ai/wiki/lenovo), Pegatron, QCT, Wistron |
| Container solution providers | [Anduril](https://aiwiki.ai/wiki/anduril), Armada, [Crusoe](https://aiwiki.ai/wiki/crusoe), DayOne, Delta, Eaton, Schneider Electric, Vertiv |

The technical blog describes the division of labor: OEMs and ODMs integrate the computing systems; mechanical, electrical, plumbing and infrastructure specialists adapt power and cooling to the site; modular solution providers manufacture the enclosures and integrate subsystems; and site partners prepare the location, install and support commissioning. NVIDIA says the model "expands the modular data center market rather than replacing existing offerings," and that partners can still sell lower levels of integration, such as modular infrastructure that customers populate on site.[2]

Two partners published their own announcements on October 21, 2026:

- **Dell Technologies** described a Dell Modular Data Center with NVIDIA DSX POD, a containerized Dell AI Factory that "will include compute, storage, networking, rack-level power and liquid cooling" and connects to prepared power. The same release listed the "Dell Modular Data Center with NVIDIA DSX Reference Design" as available in early 2027.[16]
- **Crusoe** said its Crusoe Spark modular data center platform would support the DSX POD reference design for Vera Rubin NVL72 deployments, and unveiled Crusoe Spark Liquid, which it called its sixth-generation modular AI factory. Crusoe said it has built and operated modular AI factories since 2020 and plans to mass-produce Spark units at its new 352,000-square-foot Crusoe Spark Factory in Brighton, Colorado. Bryan Schramm, Crusoe's head of Spark, said the company "can deliver productive AI capacity in as little as three months."[15]

## Reception

Coverage focused on the gap between factory speed and site readiness. IDC analyst Ashish Nadkarni told Data Center Knowledge that DSX POD could appeal for sovereign AI deployments that need physical isolation, smaller deployments built for specific needs, and edge AI environments.[7] Futurum Group called the engineering "impressive" and said DSX POD addresses a real cost, the months expensive GPUs sit idle while racks are cabled and commissioned on site, but argued that grid power, not deployment speed, is the binding constraint on AI factories in Europe. It also observed that moving integration to manufacturers shifts both value and integration risk to them, since a POD that fails acceptance testing becomes the supplier's problem rather than the customer's.[11] ServeTheHome described the European launch as no coincidence given NVIDIA's push for [sovereign AI](https://aiwiki.ai/wiki/sovereign_ai) on the continent, while noting that DSX PODs will be available worldwide.[9] CNBC framed the product as NVIDIA's latest move up the stack from chips toward complete systems.[6]

## Relationship to other NVIDIA designs

| Name | Scope |
|---|---|
| [NVIDIA DSX](https://aiwiki.ai/wiki/nvidia_dsx) | NVIDIA's overall framework for co-designed AI factory infrastructure, which customers and partners can implement in different physical forms[2] |
| DSX POD | A standardized, factory-built implementation of DSX, made of container types with defined interfaces[2] |
| Vera Rubin NVL72 | The rack-scale compute system installed in the POD; 72 Rubin GPUs and 36 Vera CPUs per rack[17] |
| Vera Rubin platform | Described by NVIDIA as a multi-rack "POD-scale system" that brings together five purpose-built rack-scale systems; a separate concept from the DSX POD container design[17] |
| [NVIDIA DGX SuperPOD](https://aiwiki.ai/wiki/nvidia_dgx_superpod) | A reference architecture for large DGX clusters, not a shipped container[2][9] |

## See also

- [NVIDIA DSX](https://aiwiki.ai/wiki/nvidia_dsx)
- [NVIDIA DSX MaxLPS](https://aiwiki.ai/wiki/nvidia_dsx_maxlps)
- [NVIDIA Vera Rubin](https://aiwiki.ai/wiki/nvidia_vera_rubin)
- [NVIDIA DGX SuperPOD](https://aiwiki.ai/wiki/nvidia_dgx_superpod)
- [Crusoe](https://aiwiki.ai/wiki/crusoe)
- [Data center](https://aiwiki.ai/wiki/data_center)
- [AI infrastructure](https://aiwiki.ai/wiki/ai_infrastructure)

## References

1. NVIDIA Newsroom. "NVIDIA Introduces DSX POD AI Factories to Accelerate Europe's Infrastructure Deployment." October 21, 2026. https://nvidianews.nvidia.com/news/nvidia-introduces-dsx-pod-ai-factories-to-accelerate-europes-infrastructure-deployment
2. Premal Savla. "Achieve Productive AI Capacity Faster and More Predictably with NVIDIA DSX POD." NVIDIA Technical Blog, October 21, 2026. https://developer.nvidia.com/blog/achieve-productive-ai-capacity-faster-and-more-predictably-with-nvidia-dsx-pod/
3. NVIDIA. "DSX POD Reference Design for AI Factory Infrastructure." Accessed October 25, 2026. https://www.nvidia.com/en-us/data-center/products/dsx/pod/
4. NVIDIA Blog. "GTC Berlin 2026: Live Updates From NVIDIA CEO Jensen Huang's Keynote." October 21, 2026. https://blogs.nvidia.com/blog/gtc-berlin-2026-keynote/
5. NVIDIA AI Infrastructure (@NVIDIAAIInfra). "Introducing NVIDIA DSX POD, a reference design for ecosystem partners to build factory-integrated containers..." X, October 22, 2026. https://x.com/NVIDIAAIInfra/status/2113405044536021422
6. Kif Leswing. "Nvidia announces plug and play Vera Rubin racks that can be shipped on a truck." CNBC, October 21, 2026. https://www.cnbc.com/2026/10/21/nvidia-announces-vera-rubin-racks-that-can-be-shipped-on-a-truck-.html
7. Wylie Wong. "Nvidia Unveils DSX Pod for Faster AI Data Center Builds." Data Center Knowledge, October 21, 2026. https://www.datacenterknowledge.com/ai-data-centers/nvidia-unveils-dsx-pod-for-faster-ai-data-center-builds
8. DCD. "Nvidia launches DSX POD, claims the Vera-Rubin reference architecture will make AI data centers more water and energy efficient." October 21, 2026. https://www.datacenterdynamics.com/en/news/nvidia-launches-dsx-pod-claims-the-vera-rubin-reference-architecture-will-make-ai-data-centers-more-water-and-energy-efficient/
9. Ryan Smith. "NVIDIA GTC Berlin 2026 NVIDIA Intros DSX PODs and More." ServeTheHome, October 21, 2026. https://www.servethehome.com/?p=106103
10. Joe Fay. "Huang shrugs off protestors to kick off GTC Berlin with DSX Pods, Robots." The Next Platform, October 21, 2026. https://nextplatform.com/control/2026/10/21/huang-shrugs-off-protestors-to-kick-off-gtc-berlin-with-dsx-pods-robots/5304447
11. Nick Patience. "At GTC Berlin, NVIDIA Puts the AI Factory in an Actual Container." Futurum Group, October 22, 2026. https://futurumgroup.com/insights/at-gtc-berlin-nvidia-puts-the-ai-factory-in-an-actual-container/
12. Larry Dignan. "Nvidia launches DSX POD, aims for modular, preintegrated AI factories." Constellation Research, October 21, 2026. https://www.constellationr.com/insights/news/nvidia-launches-dsx-pod-aims-modular-preintegrated-ai-factories
13. Jim Carroll. "NVIDIA's Factory-Built DSX POD Targets 14-Day Deployment." Converge Digest, October 21, 2026. https://convergedigest.com/nvidia-dsx-pod-vera-rubin-ai-factory-14-day-deployment/
14. The Singju Post. "Transcript of NVIDIA GTC Berlin 2026: Jensen Huang Full Keynote." Keynote delivered October 21, 2026. https://singjupost.com/transcript-of-nvidia-gtc-berlin-2026-jensen-huang-full-keynote/
15. Crusoe. "Crusoe Announces Support for NVIDIA DSX POD, Bringing Field-Proven Crusoe Spark Platform to the Expanding Modular Ecosystem." October 21, 2026. https://www.crusoe.ai/resources/newsroom/crusoe-nvidia-dsx-modular-standardization
16. Dell Technologies. "Dell Technologies Advances Enterprise Agentic AI from the Deskside to Data Center with NVIDIA." October 21, 2026. https://investors.delltechnologies.com/news-events/press-releases/detail/1659/dell-technologies-advances-enterprise-agentic-ai-from-the-deskside-to-data-center-with-nvidia
17. NVIDIA. "NVIDIA Vera Rubin Platform." Accessed October 25, 2026. https://www.nvidia.com/en-us/data-center/technologies/rubin/
18. Lucy Sherriff. "Sun touts boxed-up data centres." The Register, October 17, 2006. https://www.theregister.com/2006/10/17/sun_data/
19. TechCentral.ie. "Making sense of pre-fab data centre options." July 17, 2015. https://www.techcentral.ie/making-sense-of-pre-fab-data-centre-options