Nvidia used its GTC conference in Berlin this week to roll out a set of infrastructure, security, robotics, and networking announcements aimed at the next phase of artificial intelligence deployment. Although the event had a strong European focus, the implications extend well beyond the region. The company’s message centered on the systems that surround AI chips: how AI factories are built, how they are cooled and powered, how autonomous agents are governed, how robots are trained, and how telecom networks can support AI workloads.

The announcements reflect a broader shift in the AI buildout. The first wave focused on acquiring accelerators and standing up clusters. The next wave is about industrializing the process: shortening construction timelines, reducing energy and water consumption, securing AI agents that act on their own, and connecting AI infrastructure to physical systems such as robots and wireless networks. Nvidia used the Berlin stage to present a more integrated view of that stack.

Key facts at a glance

• Nvidia introduced DSX POD, a factory-built modular data center approach designed to cut AI infrastructure deployment from months to weeks.
• The DSX reference architecture combines closed-loop liquid cooling, dry coolers, and power-management software.
• DSX MaxLPS software can produce the same number of AI tokens using up to 30 percent less power, according to Nvidia.
• A 2-gigawatt deployment using dry cooling could save 54 million liters of cooling water per day compared with conventional cooling towers, the company estimates.
• OpenShell Reference Architecture combines Vera Rubin compute, Vera CPUs, BlueField-4 DPUs, and security software to build, monitor, and constrain autonomous AI agents.
• Nvidia expanded its Isaac robotics platform and highlighted Agile Robots’ Agile ONE SN humanoid reference platform, powered by Jetson Thor with 128GB of unified memory and planned for early 2027.
• Nvidia and Nokia expanded telecom work on AI radio access networks, with T-Mobile completing what the companies called the first AI-RAN call on a commercial network and BT, Deutsche Telekom, and Elisa announcing pilots.

AI factories move from construction projects to factory-built systems

Nvidia introduced the DSX POD, a factory-built, modular data center approach designed to accelerate AI infrastructure deployment. Instead of treating every AI data center as a one-off construction project that must be assembled from scratch, the DSX POD concept packages compute and supporting infrastructure into integrated systems. The goal is to bring manufacturing discipline to AI capacity: standardized modules can be built in a factory, shipped, and installed more quickly than fully bespoke facilities.

Nvidia is positioning DSX POD as a turnkey solution. The company says the approach could reduce deployment timelines from months to weeks. That claim matters because the economics of AI infrastructure are increasingly tied to speed. GPUs and other accelerators only create value when they are running production workloads. Every month a facility spends in construction, commissioning, or integration is a month without revenue-generating AI output.

Nvidia CEO Jensen Huang framed the issue in terms of token production. “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 underlying pressure: AI capacity is being planned at a scale where delays are costly, and operators are looking for repeatable designs rather than bespoke builds.

The modular approach also reflects lessons from the broader data center industry. Hyperscalers and large enterprises have increasingly turned to prefabricated power, cooling, and IT modules to reduce site labor, improve quality control, and cope with long lead times for components. AI factories add new constraints because high-density accelerators require sophisticated cooling and power distribution. Packaging those elements into a repeatable system could make it easier to deploy capacity in more locations, including regions where construction labor and engineering expertise are limited.

Power and water efficiency become design constraints

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 is a significant claim because power availability is one of the main bottlenecks for AI expansion. Grid connections, utility upgrades, and local permitting can take years, and many utilities are already managing queues of large load requests from data center operators.

Water use is the other constraint. 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. The figure illustrates why cooling design has moved from a back-end engineering detail to a board-level concern. Communities and regulators are paying closer attention to the water footprint of large computing facilities, especially in regions facing drought or water stress. Dry coolers can reduce dependence on evaporative cooling, while closed-loop liquid cooling can improve heat capture and reuse.

The combination of liquid cooling and power-management software is also a response to rising rack densities. Air cooling has limits, and AI workloads are pushing those limits. Liquid cooling can remove heat more efficiently from dense accelerator racks, but it introduces new plumbing, maintenance, and reliability considerations. By packaging cooling and power management into a reference architecture, Nvidia is trying to make these advanced designs easier to deploy at scale. The intended outcome is not just lower energy bills but also more predictable operations and a smaller resource footprint per unit of AI output.

OpenShell aims to constrain autonomous AI agents

Nvidia announced its OpenShell Reference Architecture, combining Vera Rubin compute, Vera CPUs, BlueField-4 data processing units, and security software to build, monitor, and constrain AI agents. The system is aimed at a fast-emerging problem: AI agents that can execute tasks across enterprise infrastructure and edge systems need guardrails that operate at machine speed. Traditional security models built around human users and static applications may not be sufficient when software agents can chain actions, call tools, and access data across multiple systems.

The OpenShell Reference 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. That means policy enforcement must be embedded in the runtime environment rather than bolted on after deployment. The architecture is available now, which signals that Nvidia sees agent security as an immediate deployment issue rather than a future research topic.

The reference architecture also shows how Nvidia is trying to tie together its data center components. Vera Rubin compute provides the AI processing foundation, Vera CPUs handle general-purpose workloads, and BlueField-4 DPUs support infrastructure processing and security functions. By combining these elements with security software, Nvidia is proposing a blueprint for organizations that want to run agentic AI workloads with more visibility and control. The practical questions will center on how policies are authored, how agent identities are managed, and how audit trails are maintained when agents act across hybrid and edge environments.

Security is becoming a gating factor for agentic AI. Enterprises are interested in agents that can automate workflows, but they are wary of giving autonomous software broad access to sensitive systems. A reference architecture that emphasizes monitoring and constraint addresses that concern at the infrastructure layer. It does not eliminate the need for governance, but it provides a technical foundation for enforcing limits on what agents can do and where they can go.

Robotics platform moves closer to commercial deployment

Nvidia expanded its Isaac robotics platform with new simulation and robot learning tools, along with a reference humanoid, datasets, and runtime, plus updates to its robot learning framework. Together, these technologies are intended to give developers an integrated path from data to deployment. Simulation is especially important in robotics because physical testing is expensive and slow. Developers can train and validate robot behaviors in virtual environments before moving them to real machines, reducing risk and accelerating iteration.

The reference humanoid and datasets are meant to lower the barrier to entry for companies building general-purpose robots. Humanoid robots are complex systems that combine perception, planning, control, and actuation. A reference platform can provide a starting point for hardware and software integration, while datasets and learning frameworks help developers train models more efficiently. Nvidia’s broader ambition is to make robotics development more like AI software development: iterative, simulation-driven, and supported by reusable tools.

Germany’s Agile Robots introduced Agile ONE SN, a humanoid reference platform built on Nvidia Isaac GR00T technology and powered by Jetson Thor. The platform has 128GB of unified memory and is scheduled to become available in early 2027. The memory specification is notable because humanoid robots must run multiple AI models on-device, including perception, language understanding, and motion control. Unified memory can simplify data movement between those workloads, which is important for real-time operation in physical environments.

The robotics announcements fit a larger pattern at the conference: AI is moving from data centers into physical systems. The same infrastructure that trains models is increasingly connected to the systems that act on those models. Simulation tools, reference hardware, and runtime software are the connective tissue. For robotics developers, the appeal is a shorter path from prototype to commercial deployment. For Nvidia, the opportunity is to extend its AI computing platform into a new class of machines.

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. BT, Deutsche Telekom, and Elisa announced pilots using Nokia software accelerated by Nvidia technology. AI-RAN is an effort to bring AI computing into radio access networks so that operators can allocate spectrum and network resources more efficiently, support new services, and potentially run AI inference closer to users.

Telecom networks are a natural testbed for AI infrastructure because they combine real-time constraints, distributed sites, and high reliability requirements. A radio access network must make split-second decisions about how to use spectrum and serve users. Adding AI to that environment requires careful integration with existing network functions. The pilots announced at GTC Berlin are early, but they show that operators are exploring how AI can improve network performance and create new edge computing opportunities.

The T-Mobile call is a milestone because it was completed on a commercial network rather than in a lab. The operator pilots with BT, Deutsche Telekom, and Elisa will test how Nokia software accelerated by Nvidia technology performs in real deployment conditions. Those trials will be watched for evidence that AI-RAN can deliver efficiency gains without compromising reliability. They also illustrate how AI infrastructure is spreading beyond centralized data centers into the network edge, where latency, power, and security constraints are different from those in a hyperscale facility.

Source:Network World News