News
Nvidia at GTC Berlin: Containerized AI Factories in as Little as 14 Days and a Security Architecture for Agents
At GTC Berlin, Nvidia unveiled DSX POD, prefabricated AI factories built on Vera Rubin, and the OpenShell Reference Architecture for secure agents. The company says deploying an AI factory at a prepared site can drop from six months to as little as 14 days.
Contents
At the GTC Berlin conference on October 21, Nvidia announced two major projects for European AI infrastructure: DSX POD, plug-and-play AI factories built in modular containers, and the OpenShell Reference Architecture, a framework for running AI agents under security policies. Both announcements are described in the company's official newsroom.
DSX POD Ships AI Factories in Containers
DSX POD is a reference design for complete AI factories on the Vera Rubin platform. Partners build them in modular containers, then integrate and validate them before delivery. The finished unit arrives at a site with power, cooling and networking already prepared.
According to Nvidia, compute, networking, cabling, rack-level power and liquid cooling are installed at the manufacturer's facility rather than on site. Partners also complete rack-to-rack cabling, piping and functional testing before shipment. The company argues that factory testing reduces the risk of component delays and errors in onsite construction work.
Nvidia says that at a prepared site, deployment time falls from six months to as little as 14 days. Worldwide deliveries are due to start in early 2027. The named partners include Bull, Dell Technologies, HPE, Lenovo, Schneider Electric, Vertiv, Foxconn (Ingrasys), GIGABYTE and Wistron.
“An AI factory creates value when it starts producing tokens — not when construction is complete,” - Jensen Huang, founder and CEO of NVIDIA
OpenShell for Agents
The second project is the OpenShell Reference Architecture, a full architecture for building, containing, observing and enforcing policies on AI agents in enterprises and at the edge. It builds on the open source OpenShell runtime.
On the hardware side, Vera Rubin NVL72 handles inference, Vera CPUs run the agent loops and tools, and BlueField-4 DPUs with Sentry software provide oversight. Sentry monitors an agent's behavior from outside its environment and enforces policies in silicon. According to Nvidia, an agent that tries to move beyond its software boundary is isolated and stopped within milliseconds. Huang's summary: agents think on Rubin, act on Vera and are governed on BlueField.
The named partners include model builders (among them Anthropic, Google, Meta and OpenAI), security vendors (among them CrowdStrike, Palo Alto Networks and Microsoft) and software companies such as SAP, Siemens and IBM. Banks BBVA and Santander are described as only exploring the framework. ServiceNow announced general availability of ServiceNow Arc, an agent governed by Control Tower and enforced by OpenShell.
Caveats and Vendor Figures
All performance figures come from Nvidia. That includes the claimed 30x higher throughput per megawatt for agentic workloads on Vera Rubin and 2x the performance of x86 processors for Vera CPUs. The announcement also carries a disclaimer that some features will only be available in the future and may change at the company's discretion.
Separately, Nvidia says its DSX AI Factory platform lets companies worldwide plan, simulate and operate the power grid and the AI factory together, from the substation to the rack. The company says power is now the biggest bottleneck in infrastructure.
Robotics and Local Hardware
Nvidia also announced progress on its Isaac platform for general-purpose robotics, including a new reference humanoid.
On the local computing side, RTX Spark laptops are already available from $2,399 from ASUS, Dell, HP, Lenovo, Microsoft and MSI, while DGX Spark is aimed at developers building and scaling applications on Linux. DGX Station for Windows is due in December.
What It Means for Poland
Polish companies and data center operators struggling with grid connections and construction timelines get another argument for prefabrication. But a faster installation alone does not solve the problem of power and connections, since DSX POD assumes a ready site with power and cooling. Real timelines and prices will only become clear once deliveries begin in 2027.
