NVIDIA's DSX POD Ships AI Factories from the Factory in Containers — Deployment Cut from Six Months to as Little as 14 Days, with Power and Liquid Cooling Built In; Infineon Brings GaN to 800 VDC Intermediate Bus Conversion

On October 21, 2026, at GTC Berlin in Germany, NVIDIA announced "DSX POD," a reference design for shipping AI factories from the factory in containers. Partners assemble a complete Vera Rubin AI factory inside modular containers and integrate and validate it before shipment. Once connected to a prepared site, it can begin operating, and NVIDIA says it cuts deployment time from six months to as little as 14 days.

On the same day, Infineon announced it is advancing its collaboration with NVIDIA on 800 VDC power architecture, using GaN in the high-voltage intermediate bus converter (IBC) inside the rack. The assumptions behind AI data center power are shifting both in how buildings are put together and in power conversion inside the rack.

Moving site work to the factory

With DSX POD, integration and installation of compute, networking, cabling, rack-level power and liquid cooling move to the manufacturing plant. Partners complete inter-rack cabling, electrical wiring, plumbing and functional validation before shipment, in parallel with site preparation.

The initial focus is large-scale, newly built liquid-cooled AI infrastructure, such as that built by NVIDIA Cloud Partners (NCPs). NVIDIA says conventional construction and other modular approaches remain valid depending on the use case. The product page also shows the use of DSX MaxLPS, power control software that maximizes throughput within a power limit.

Seventeen companies are taking part, with worldwide supply starting in early 2027.

*Grouping by area is ours.

Four major power and cooling vendors join container-level turnkey delivery in the same framework as server makers. Delta has separately announced a modular data center including 800 VDC in-row power.

GaN in the in-rack step-down stage

Infineon's announcement focuses on the intermediate bus converter (IBC) within the 800 VDC power chain. The IBC is the stage that steps the main bus voltage down to lower voltages for accelerator boards; later stages generate the AI processor core voltage and point-of-load voltages.

Infineon positions 800 VDC as an additional distribution option alongside existing AC and low-voltage DC, and says GaN's low losses limit heat as power per rack approaches the megawatt level. The company has also announced a 27 kW three-phase AI server power supply for the rack-entry side.

What it means for facility and power design teams

• Time savings depend on site readiness: DSX POD shortens on-site installation. Without secured power capacity, a grid connection and cooling water supply, the 14-day figure does not apply. The need to settle land and power intake plans first is unchanged
• The unit of power and cooling procurement changes: rack power and liquid cooling equipment will increasingly be delivered as containers integrated with servers. Specifications and maintenance contracts need to be checked at container level, not for individual equipment
• The 800 VDC discussion moves to the on-board step-down stage: the choice of 800 V to 50 V/12 V/6 V feeds directly into server board power design. The conversion ratio chosen changes the configuration of downstream voltage regulators (for power delivery right before the xPU, see Navitas's acquisition of Claros)
• What has not been disclosed: power capacity and pricing per DSX POD container, and the efficiency and volume production timing of Infineon's GaN IBC

Referenced Fact Cards