Back to the stories

NVIDIA introduces DSX AI Factory platform and DSX POD reference design for liquid-cooled AI infrastructure

Score 9.7,

AI infrastructure and compute

You can now plug a mostly pre-built, liquid-cooled AI factory into a site and be running high-end training and inference within weeks.

Behind that change is NVIDIA's new DSX AI Factory platform and a modular reference called DSX POD: a packaged design that coordinates compute, networking, storage, software, power and cooling as one system for Vera Rubin deployments.

NVIDIA says DSX designs cut cooling-water use and improve power efficiency. The DSX POD is sold as a pre-integrated unit that can be deployed on site in about 14 days after utilities are available, and partners including Nscale and Crusoe are already lining up to build on the design.

Why this matters: building big AI capability used to mean long construction schedules, custom engineering and a lot of guesswork about cooling and power. DSX turns much of that into an off-the-shelf product, so operators can add capacity faster, with less trial and error and lower ongoing energy cost.

How it works, simply: think of DSX as a factory in a box. Most of the complex plumbing, cabling and systems arrive assembled, and liquid cooling runs chilled fluid directly to the racks instead of relying on warm air. That focused approach lets a site pack more GPUs into the same space and move heat away more efficiently.

What changes now: cloud operators and regional providers in Europe can scale dense AI capacity faster and with better energy economics. Early tests from Nscale show higher GPU density and better inference throughput with DSX-based hardware. At the same time, this is not a desktop trick: sites still need enterprise-grade power and cooling hookups.

What's next: the real test will be independent benchmarks, broader GA rollouts and whether independent operators can turn this reference design into a cheaper, faster alternative to renting frontier compute. Will modular AI factories shift the balance from cloud-only to regional ownership? We should know within months.