NVIDIA DSX POD Turns AI Data Centers Into Factory-Built Products, Cutting On-Site Deployment to 14 Days

On 21 October 2026, at GTC Berlin, NVIDIA unveiled its DSX POD prefabricated AI data center reference design. According to NVIDIA, on-site deployment time can be shortened from roughly six months to as little as fourteen days, provided external infrastructure is already in place. The more significant point is not speed alone: AI data centers are moving from traditional construction projects toward standardized, industrialized products.

1. Moving the AI Data Center Into the Factory

Conventional AI data centers require extensive on-site work — rack installation, network cabling, power connections, liquid-cooling piping and system commissioning. Multi-discipline coordination, complex interfaces and long commissioning cycles constrain delivery efficiency. DSX POD reverses the logic: system integration and validation happen in the factory, while the site handles module installation, external connections and joint commissioning. Delivery follows three stages: factory integration of GPU racks, network, power distribution, liquid cooling, piping and control systems with standardized assembly and testing; module transport of validated integrated modules to site; and on-site commissioning covering external power, cooling and network connections plus system re-verification and start-up.

The core idea is not simply packing equipment into containers, but consolidating scattered engineering steps into a standardized industrial manufacturing process.

2. Key Technical Changes

DSX POD is built on NVIDIA's next-generation Vera Rubin computing platform and involves four major shifts. First, 800V DC power input: core modules adopt a standardized 800V DC input design, reducing power-interface differences between projects, although upstream generation, transformation, rectification and regional electrical infrastructure still require project-specific design. Second, prefabricated liquid cooling: piping, valves, monitoring and related equipment are integrated, flushed, pressure-tested and validated in the factory, cutting on-site construction and commissioning. Third, transport becomes part of the design: weight, center of gravity, lifting, transport vibration and equipment fixation must be considered from the design stage, and road, rail and sea freight limits directly determine module size, integration level and engineering cost. Fourth, software and hardware operate together: the NVIDIA DSX software stack combines power management, load scheduling and real-time monitoring to raise compute utilization within a given power envelope.

3. Who Builds These AI Factories?

NVIDIA provides the reference architecture, technical specifications and system validation foundation; actual manufacturing and engineering delivery remain with partners. Published partners include Dell, HPE, Lenovo, Foxconn and QCT for servers and system integration; Schneider Electric, Eaton and Delta for power and infrastructure; Vertiv and Armada for modular infrastructure; and Crusoe and DayOne for data center development and delivery. NVIDIA plans to begin global DSX POD deliveries with partners in early 2027, although the model remains at an early commercial stage, with cost, delivery capability and long-term reliability still to be verified.

4. Implications for US and Southeast Asian AIDC Development

Three directions deserve attention. First, construction speed is converting into financial competitiveness: bringing compute capacity online one month earlier means one month of earlier token revenue and operating cash flow, and less idle time for expensive GPUs, so investment models should weigh time-to-first-revenue, utilization and capital occupation alongside cost per MW. Second, energy infrastructure remains the developer's core competence: DSX POD does not solve land, power, cooling, network or permitting, and behind-the-meter gas generation projects in the United States still require independent development of gas supply, gensets, distribution, external cooling and approvals. Third, Chinese enterprises gain new supply chain opportunities as integration shifts to the factory — modular steel structures and equipment integration, low and medium-voltage distribution and power conversion, liquid-cooling equipment and piping, factory prefabrication, testing and quality control, international transport, lifting and site installation, and overseas infrastructure delivery.

5. From Building Delivery to Compute Delivery

NVIDIA is pushing AI data centers from a construction-driven delivery model toward an industrialized, compute-system-driven one. Traditional design institutes, EPC contractors and site construction are not being replaced, but their scope and value distribution will change. One caveat must be explicit: fourteen days is the fastest deployment time once external infrastructure is ready, not the time from land acquisition to a fully commissioned data center. The real development cycle still depends on land, power, permits, manufacturing, transport and commissioning. Yesterday the metrics were floor area, construction cost and completion date; tomorrow they may be effective tokens generated per MW of investment, and how long it takes to turn capital into compute revenue.