AI factories are starting to arrive as pods. NVIDIA's DSX POD is a Vera Rubin reference design that partners build and deliver. NVIDIA claims deployment in as little as 14 days at a prepared site, with partner deliveries targeted for early 2027.
The catch is in the words prepared site. Power, heat rejection, connectivity, permits and commissioning still have to be ready before the pod arrives.
What does modular include? It is not one thing:
- Reference design: a validated blueprint for partners to build.
- Integrated modules: IT, cooling and power assembled in the factory.
- Compute module: racks integrated with cooling and power distribution.
- Power infrastructure: electrical skids and eHouses.
- Modular data hall: prefabricated space and supporting systems.
- Building blocks: components configured into a project solution.
Vendors can span several categories, and how much IT is included varies.

The providers. HPE's AI Mod POD supports AI, HPC and private cloud AI with direct liquid and air cooling: the DC8 configuration takes up to 0.8 MW of IT across 8 racks and DC16 up to 1.6 MW across 16, with a PUE target of 1.14. Around it, Schneider Electric (RapidScale), Flex (Axiom AI Modules), Armada (Leviathan and Orion), Eaton with Flexnode (prefabricated data halls), Siemens with Delta (skids and eHouses), Delta (prefabricated AI MDC), GIGABYTE (GADU), LG CNS (PMDC) and Crusoe (Spark Liquid) each cover a different part of the stack. Capacity and IT inclusion depend on configuration.
From factory floor to first token: how fast can you get there?
