A new infrastructure approach developed by Nvidia and Vertiv shrinks the timeline for bringing AI factories to productive capacity from 120 days to 48 days. This method, which combines the Vera Rubin DSX reference design with prefabricated infrastructure, can increase cumulative AI token output by up to 30% in the first year. CEO Jensen Huang warned that delay costs are extreme, noting that a three month setback for a 1 GW data center—costing $50 billion to build—can result in $10 billion in lost annual revenue. Polar Data Centers is applying this approach to the initial 41 MW phase of its facility in Norway.
The efficiency gains build on the DSX POD modular units introduced at GTC Berlin, which cut on-site installation at prepared sites from six months to as little as 14 days. These containerized systems, which integrate compute, networking, and liquid cooling, are slated for worldwide availability in early 2027 through 17 partners including Dell and Foxconn. The DSX platform also reduces power use by up to 30% per token via MaxLPS and cuts cooling water use to near zero at a 2 GW scale.