Nvidia used its Berlin GTC event to roll out a set of infrastructure and software announcements aimed at the physical and operational bottlenecks now shaping artificial intelligence deployment. The company introduced a factory-built modular data center concept called DSX POD, outlined cooling and power-management designs intended to cut electricity and water use, extended its security stack to autonomous AI agents, expanded robotics tools, and reported new telecom trials for AI-powered radio access networks. Although the event had a European focus, the announcements address problems that data center operators face in every region: how to build AI capacity faster, how to keep it within power and water limits, and how to govern agents that act on enterprise systems.

Key facts

• Nvidia announced DSX POD, a factory-built modular data center approach that packages compute and supporting infrastructure into integrated systems rather than requiring each deployment to be assembled from scratch.
• The company positioned DSX POD as a turnkey model that could reduce deployment timelines from months to weeks.
• Nvidia’s DSX reference architecture combines closed-loop liquid cooling, dry coolers and power-management software.
• DSX MaxLPS software is designed to generate the same number of AI tokens using up to 30 percent less power.
• Nvidia estimates that a 2-gigawatt deployment using dry cooling could save 54 million liters of cooling water per day compared with conventional cooling-tower systems.
• The OpenShell Reference Architecture combines Vera Rubin compute, Vera CPUs, BlueField-4 data processing units and security software to build, monitor and constrain AI agents.
• Nvidia expanded its Isaac robotics platform with simulation and robot-learning tools, a reference humanoid, datasets and runtime, plus updates to its robot-learning framework.
• Germany’s Agile Robots introduced Agile ONE SN, a humanoid reference platform built on Isaac GR00T technology and powered by Jetson Thor, with 128GB of unified memory and availability scheduled for early 2027.
• Nvidia and Nokia expanded work with telecom operators on AI radio access networks, with T-Mobile completing what the companies described as the first AI-RAN call on a commercial network.

Modular AI factories target the construction bottleneck

The centerpiece of the Berlin announcements was DSX POD, a modular approach that treats AI data center capacity as a factory-built product rather than a one-off construction project. In the conventional model, operators design a facility, procure equipment, coordinate mechanical and electrical contractors, and then integrate compute, networking and cooling on site. That process can take months and can leave expensive accelerators waiting for buildings and power systems to catch up. Nvidia’s pitch is that integrated systems built in a controlled manufacturing environment can shorten that path and reduce variability.

Nvidia is positioning DSX POD as a turnkey solution for customers that need to add AI capacity quickly. The concept packages compute and supporting infrastructure into integrated systems, which could make deployments more repeatable and easier to replicate across sites. That matters for cloud providers, sovereign AI projects and enterprises that are trying to convert capital budgets into production capacity before hardware generations change again. The company says the approach could reduce deployment timelines from months to weeks.

Jensen Huang, founder and CEO of Nvidia, framed the issue in terms of production rather than construction. “An AI factory creates value when it starts producing tokens — not when construction is complete,” Huang said in a statement. “With Nvidia DSX POD, we’re bringing factory-built precision and speed to AI infrastructure construction — reducing deployment time from months to weeks so customers can scale capacity and bring revenue-generating production online faster.” The quote captures the economic logic behind modular designs: in AI infrastructure, delays are not just schedule problems. They are revenue problems, because expensive compute sits idle while construction and integration work proceed.

Power and water become first-order design constraints

Nvidia’s infrastructure story is increasingly inseparable from energy policy. AI factories are dense with accelerators, networking and storage, and the power required to run them is only part of the challenge. Cooling systems add electrical load and, in many designs, consume large volumes of water. Utilities and local governments are scrutinizing both. That has pushed liquid cooling, closed-loop designs and dry cooling from niche options into mainstream planning discussions for high-density AI facilities.

Nvidia’s DSX reference architecture combines closed-loop liquid cooling, dry coolers and power-management software. The company says its DSX MaxLPS software can generate the same number of AI tokens using up to 30 percent less power. That claim is significant because efficiency gains compound at scale. A percentage improvement in power per token can translate into lower operating costs, more capacity within an existing power envelope, or the ability to deploy in locations where grid connections are constrained.

Water use is the other side of the resource equation. Nvidia estimates that a 2-gigawatt deployment using its dry-cooling approach could save 54 million liters of cooling water per day compared with conventional cooling-tower systems. That estimate will be tested in real deployments, but it signals how vendors are trying to address environmental permitting and community concerns alongside performance. For operators, the design trade-offs are not abstract. Cooling choices affect site selection, utility negotiations, maintenance practices and long-term operating costs. Nvidia is trying to make those choices part of a reference architecture rather than a bespoke engineering exercise for every facility.

OpenShell brings security controls to autonomous agents

The software side of the Berlin event focused on a fast-emerging problem: how to let AI agents act on enterprise systems without giving them unchecked authority. Agents are moving from answering questions to executing tasks, calling APIs, retrieving data and interacting with other software. That shift creates value, but it also creates new attack surfaces. An agent with access to calendars, code repositories, procurement systems or customer records can become a path for data leakage, prompt injection or unintended actions.

Nvidia announced its OpenShell Reference Architecture, which combines Vera Rubin compute, Vera CPUs, BlueField-4 data processing units and security software to build, monitor and constrain AI agents. The architecture builds on Nvidia OpenShell secure runtime software and extends it into full-stack, industrial-scale infrastructure for agent development and operations. The design is intended to enforce security policies as agents execute tasks across enterprise infrastructure and edge systems. The reference architecture is available now.

The timing reflects a broader industry shift. Enterprises are experimenting with agents for software development, customer support, security operations and back-office workflows, but many are hesitant to grant broad permissions. A secure runtime can provide guardrails: defining what an agent can access, logging its actions, isolating workloads and applying policy at the infrastructure layer. Nvidia’s approach ties that software control to specific hardware components, including DPUs that can handle networking and security functions. Whether customers adopt a full reference architecture or mix components, the announcement signals that agent security is becoming an infrastructure category rather than an application add-on.

Robotics moves closer to commercial deployment

Nvidia also expanded its Isaac robotics platform with new simulation and robot-learning tools, a reference humanoid, datasets and runtime, along with updates to its robot-learning framework. The company says these technologies give developers an integrated path from data to deployment to bring general-purpose robots to the real world faster. That path matters because robotics development has traditionally been fragmented: simulation, model training, hardware integration and runtime deployment often come from different vendors and require custom glue work.

Simulation is especially important for humanoids and other general-purpose robots. Training in the physical world is slow, expensive and sometimes dangerous. Simulated environments can generate synthetic data, test control policies and expose robots to scenarios that would be difficult to stage in a lab. Nvidia’s broader strategy is to connect those simulated environments to real hardware through common software and compute platforms, reducing the iteration cycle between training and deployment.

Germany’s Agile Robots introduced Agile ONE SN, a humanoid reference platform built on Nvidia’s Isaac GR00T technology and powered by Jetson Thor. It has 128GB of unified memory and is scheduled to become available in early 2027. A reference humanoid does not mean every customer will build the same robot, but it gives developers a common baseline for testing perception, manipulation and learning workflows. For the robotics ecosystem, that can lower the cost of experimentation and make it easier to compare results across hardware designs.

Telecom operators test AI-powered networks

Nvidia and Nokia announced expanded work with telecom operators on AI radio access networks. T-Mobile completed what the companies described as the first AI-RAN call on a commercial network, while BT, Deutsche Telekom and Elisa announced pilots using Nokia software accelerated by Nvidia technology. AI-RAN is an attempt to bring AI compute into radio access networks, potentially improving spectrum efficiency, energy management and network optimization.

Telecom networks are attractive candidates for AI because they generate enormous amounts of operational data and must make real-time decisions about traffic, interference and power. But they also have strict reliability requirements. Operators cannot afford to experiment with core connectivity in ways that compromise service. That is why the first commercial call and the operator pilots matter: they move AI-RAN from conceptual architecture to controlled field testing on live or near-live networks.

The Berlin GTC announcements, taken together, show Nvidia pushing beyond the accelerator chip into the surrounding systems that determine whether AI capacity can actually be built and operated. Modular factories address time to deployment. Cooling and power software address energy and water constraints. OpenShell addresses agent governance. Isaac addresses robotics development. AI-RAN addresses edge networks. The common thread is integration: Nvidia is trying to package compute, networking, software and reference designs so customers can move faster across each of those domains.

For European operators and enterprises, the Berlin setting also reflects a broader policy context. Sovereign AI, data residency and energy availability are shaping investment decisions across the region. Nvidia’s message is that AI infrastructure can be deployed in a more standardized, resource-aware way, which may make it easier for organizations to justify projects under tighter power and permitting regimes. The OpenShell reference architecture is available now, while the Agile ONE SN platform is scheduled for early 2027, and the telecom pilots will test how quickly AI-RAN moves from commercial first calls to broader network deployment.

Source:Network World News