Nvidia DSX POD puts a Vera Rubin AI factory in a container

Nvidia unveiled DSX POD at its GTC Berlin conference on October 21, 2026: a complete AI data center built inside a shipping container. Nvidia says a container arrives tested and can be running in 14 days, instead of the six months an on-site build takes. Jensen Huang also announced nearly $10 billion of commitments to Europe's technology companies.

What DSX POD is

DSX POD is a reference design, a blueprint that partner companies build and sell. It packs Vera Rubin, Nvidia's newest AI computing system, into modular containers that are integrated and checked before delivery, according to the Nvidia keynote blog.

ServeTheHome gives the details. Each container holds seven Vera Rubin NVL72 racks, plus networking, cooling and management gear. It uses a standard intermodal container size, so it ships by truck. The site still has to supply power, network links and a larger cooling system.

Nvidia will not build or sell DSX PODs itself, ServeTheHome reports. Partners including Gigabyte, HPE, Dell, Delta and QCT will build them, with delivery worldwide expected in early 2027.

"An AI factory creates value when it starts producing tokens, not when construction is complete," Huang said, as quoted by Analytics India Magazine. Tokens are the units of text a model reads and writes.

Power, water and the supply chain

Nvidia says its DSX AI factory platform produces the same AI output with up to 30% less power. It also claims cooling water use can drop by up to 100%. Analytics India Magazine attributes the power figure to a product called DSX MaxLPS. A separate tool, DSX Flex, adjusts power use to conditions on the electricity grid.

Nvidia also announced deals to expand the parts supply behind its chips:

Area	Target
Substrates (IBIDEN, Samsung Electro-Mechanics, Unimicron)	2.5x capacity over three years
Amkor interposer wafers	4x by the end of 2028
KYEC chip testing	2x by 2027
SPIL co-packaged optics	15x, rising to 30x by the end of 2028

Nvidia is backing the expansion with a $1.1 billion capacity-reservation advance, Analytics India Magazine reports. "Demand is extraordinary. Capacity is the constraint," Huang said.

Nearly $10 billion for Europe

The Europe pledge followed a week-long tour of Paris, London, Stockholm and Dublin, Nvidia says. StorageNewsletter lists some of the pieces:

• UK: $250 million of computing over five years for research into the safety of autonomous AI agents.
• Sweden: Sferical One, a shared supercomputer with 1,152 GPUs, backed by Ericsson, AstraZeneca, Saab, SEB and Wallenberg Investments.
• Nordics: 1 gigawatt of AI cloud capacity projected by the end of 2027.
• Germany: work with Deutsche Telekom on a sovereign industrial AI cloud, meaning one run under European control.

Agents and robots

Nvidia also published the OpenShell Reference Architecture for running AI agents safely. In it, Vera Rubin runs the models, Vera CPUs run the agents, and BlueField-4 chips enforce security, ServeTheHome reports. "Every single time an agent is launched, the privileges are different," Huang said, according to Nvidia's blog.

For robotics, Nvidia showed Agile ONE SN, a reference humanoid robot built on its Isaac GR00T models and expected next year.

What this means for developers

Most developers will never buy a container of Vera Rubin racks, but its effects reach anyone renting AI compute.

• Expect faster capacity, not cheaper capacity. If containers cut deployment from months to days, new GPU capacity can appear faster from 2027. Huang's own words say demand still outruns supply.
• Plan for regional AI clouds in Europe. The Nordic and German projects add options for teams that must keep data in the EU. Ask cloud providers where their Vera Rubin capacity will sit.
• Look at the agent security design now. Nvidia released the OpenShell reference software on October 21. Teams building agents can study how Nvidia separates the model, the agent and the security checks.
• Treat power and water claims as vendor figures. The 30% and 100% numbers are Nvidia's own. Ask for measured results before they shape a data center budget.