NVIDIA Introduces DSX POD AI Factory Design with Six Taiwanese Partners

Preface

NVIDIA (輝達, NVDA-US) announced on the 21st that it is introducing DSX POD, a modular design for AI factories. The global partner network includes six Taiwanese companies or business units: Ingrasys, Delta Electronics (台達電, 2308-TW), Pegatron (和碩, 4938-TW), QCT, Wistron (緯創, 3231-TW), and Gigabyte (技嘉, 2376-TW). Deliveries worldwide are expected to begin in early 2027.

The initiative focuses on a practical challenge in AI infrastructure: bringing computing capacity into operation quickly. NVIDIA says DSX POD moves much of the equipment integration and verification work from the construction site to the factory. Its stated aim is to shorten deployment timelines, allowing operators to begin producing tokens and serving workloads sooner. This article explains the design, deployment process, partner roles, and potential operational implications described in the announcement.

Lazy bag

DSX POD is a repeatable, container-based approach to building AI factories. It combines Vera Rubin AI systems with networking, power, and liquid-cooling infrastructure, with integration and validation completed before delivery. NVIDIA says this can reduce AI factory deployment time from 6 months to as little as 14 days at sites already prepared with power, cooling, and network infrastructure. The partner list includes six Taiwanese participants, while global deliveries are expected to begin in early 2027. The central idea is to shift complex integration work into the factory so customers can bring computing capacity online sooner.

Main Body

NVIDIA’s DSX POD is a modular AI factory architecture designed to package computing systems and their supporting infrastructure into repeatable units. The design brings together Vera Rubin AI systems, networking, power equipment, and liquid-cooling systems. Rather than shipping these elements separately and relying on extensive integration at the destination, partners assemble and validate the systems inside modular containers before delivery.

The announcement reflects a distinction NVIDIA makes between completing a building and creating operational value. Founder and Chief Executive Officer Jensen Huang (黃仁勳) said an AI factory begins to create value when it starts generating tokens, not simply when construction work is finished. In this context, tokens are outputs generated by AI models. The faster a facility can begin processing models and handling customer workloads, the sooner its operator can offer services and pursue revenue-generating activity.

NVIDIA presents DSX POD as a way to apply the precision and speed associated with factory manufacturing to AI infrastructure construction. It says the approach could reduce deployment from months to weeks. More specifically, the company states that the design can cut AI factory deployment time from 6 months to as little as 14 days, provided the destination site already has the necessary power, cooling, and network infrastructure in place. The broader delivery schedule for the new architecture is expected to begin worldwide in early 2027.

The site-readiness condition is important. DSX POD does not eliminate the need to prepare a location for a large computing installation. A facility must still have adequate electrical capacity, cooling provision, and network connectivity before the modular system arrives. Instead, the design aims to reduce the amount of integration and installation that must take place after delivery. With essential site infrastructure ready, the pre-integrated system can be connected and brought into use with fewer steps than a conventional build.

Under a traditional approach, computing equipment, networking components, cables, rack power systems, and liquid-cooling equipment may require extensive coordination and installation at the site. That work can involve advance planning, engineering, specialist labor, and logistics. NVIDIA says DSX POD relocates the integration and installation of these components to the factory. Partners complete the work in a controlled environment, where they can test the system before it is shipped.

Factory testing and validation are intended to address several sources of delay and cost. If equipment arrives separately, a project may be affected by late component deliveries or by errors discovered only after on-site construction begins. Completing integration in advance can reduce the likelihood of those issues arising during deployment, according to NVIDIA. It may also limit extra expenses associated with correcting installation problems at the destination. The announcement describes these as intended benefits of the approach rather than a guarantee that every deployment will avoid delays.

The design uses repeatable, container-based units that partners can manufacture and deliver at scale. The Vera Rubin system and the infrastructure required to operate it are integrated together in a dedicated container. After partners complete system integration and verification, the container is transported to the deployment location. This structure is intended to make the delivery process more standardized than assembling each installation entirely on site.

DSX POD continues the ideas behind NVIDIA’s DSX infrastructure platform, while using a different system architecture. The stated distinction is that DSX POD places the equipment integration and verification inside dedicated containers before transport. The modular format is also intended to support production across multiple partners, allowing the design to be manufactured and delivered repeatedly rather than treated as a one-off construction project.

The partner ecosystem includes companies with roles across computing systems, networking, power, cooling, and infrastructure. Taiwanese participants named in the announcement are Ingrasys, a subsidiary of Foxconn (鴻海, 2317-TW); Delta Electronics (台達電, 2308-TW); Pegatron (和碩, 4938-TW); QCT, a subsidiary of Quanta Computer (廣達, 2382-TW); Wistron (緯創, 3231-TW); and Gigabyte (技嘉, 2376-TW). The announcement identifies these companies or business units as partners in the wider DSX POD effort.

NVIDIA’s global partner list also includes Anduril, Armada, Bull, Crusoe, DayOne, Dell Technologies (戴爾科技), Eaton, HPE (慧與科技), Lenovo (聯想), Schneider Electric (施耐德電機), and Vertiv, alongside the Taiwanese participants. This broad group reflects the range of capabilities required to deliver a functioning AI factory. Computing hardware alone is not sufficient: the installation also depends on power delivery, cooling, networking, rack systems, cabling, and coordinated deployment work.

Before shipment, partners are expected to complete inter-rack cabling, wiring, piping, and functional verification. These tasks can proceed at the factory while site preparation is underway, allowing two workstreams to happen in parallel. The site team can prepare power, cooling, and network infrastructure as partners integrate and test the modular system. Once both are ready, the completed unit can be delivered and connected, reducing the time between equipment arrival and operational use.

Deployment speed can affect an AI factory’s economics because operational value depends partly on when the facility begins serving workloads. The earlier an operator can generate its first token, run model computations, and serve production workloads, the sooner it can make capacity available to users. NVIDIA’s argument is that a shorter interval between construction and operation can help operators start providing services and generating revenue sooner.

However, the announcement’s timeline is tied to a specific deployment model and prepared sites. The 14-day figure is presented as the fastest deployment time, not as a universal schedule for every project. Actual timelines may depend on site readiness, equipment availability, project coordination, and the requirements of each installation. Likewise, the early 2027 delivery expectation describes the planned start of global deliveries rather than the completion date for all deployments.

Overall, DSX POD is positioned as a manufacturing-led approach to AI infrastructure. Its defining features are modular containers, pre-shipment integration and validation, and coordination between factory production and site preparation. By shifting more work away from the deployment location, NVIDIA and its partners aim to make AI factory construction more repeatable and to shorten the path from installed equipment to useful computing capacity.

Key Insights Table