Philadelphia Live News

collapse
Home / Daily News Analysis / Nvidia GTC Berlin: Faster AI factories, lower power use, more secure agents

Nvidia GTC Berlin: Faster AI factories, lower power use, more secure agents

Oct 24, 2026  Twila Rosenbaum  73 views
Nvidia GTC Berlin: Faster AI factories, lower power use, more secure agents

Nvidia used its GTC Berlin conference this week to roll out a set of infrastructure, security, robotics and telecom announcements aimed at the companies building and operating AI systems. The event had a strong European focus, but the technologies touch global data center buildouts, enterprise AI agent deployments and next-generation mobile networks. The common thread is Nvidia's effort to turn AI infrastructure from a collection of components into repeatable, factory-built systems that can be deployed faster and operated with tighter control over power, water and security.

Key facts at a glance

  • Nvidia introduced DSX POD, a factory-built modular data center approach for AI infrastructure that the company says can cut deployment timelines from months to weeks.
  • The DSX reference architecture combines closed-loop liquid cooling, dry coolers and power-management software; DSX MaxLPS is claimed to deliver the same AI token output using up to 30% less power.
  • Nvidia estimates that a 2-gigawatt deployment with dry cooling could save 54 million liters of cooling water per day compared with conventional cooling towers.
  • The OpenShell Reference Architecture combines Vera Rubin compute, Vera CPUs, BlueField-4 DPUs and security software to build, monitor and constrain autonomous AI agents; it is available now.
  • Nvidia expanded its Isaac robotics platform with simulation and robot-learning tools, datasets, runtime and a reference humanoid; Agile Robots introduced the Agile ONE SN humanoid, built on Isaac GR00T and powered by Jetson Thor, with 128GB unified memory and availability planned for early 2027.
  • Nvidia and Nokia expanded telecom work on AI radio access networks, with T-Mobile completing what the companies described as the first AI-RAN call on a commercial network and BT, Deutsche Telekom and Elisa announcing pilots.

Modular AI factories aim to compress construction timelines

The centerpiece of the Berlin rollout was DSX POD, which packages compute and supporting infrastructure into integrated systems rather than requiring each deployment to be designed and assembled from scratch. Nvidia is positioning the design as a turnkey route to AI capacity, with the promise that deployment can be measured in weeks instead of months. That claim matters because AI data center projects are increasingly constrained less by demand than by the time it takes to secure power, build shells, install cooling and commission complex systems.

Traditional data center construction follows a sequential process: site selection, permitting, power contracts, building construction, equipment installation and integration. Each stage can introduce delays. A modular approach moves much of the integration into a factory, where repeated manufacturing processes can improve quality control and reduce on-site labor. Nvidia's pitch is that the AI facility should be treated less like a bespoke construction project and more like an industrial product. The company says the approach can reduce deployment timelines from months to weeks, allowing customers to scale capacity and bring revenue-generating production online faster.

The commercial logic is tied to when an AI factory begins producing useful output. Nvidia CEO Jensen Huang framed the issue in terms of tokens, the units of text and other data generated by AI models. "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."

That emphasis on speed reflects a broader shift in AI infrastructure. Early GPU clusters were often installed in existing facilities or purpose-built halls with long lead times. As models grow and inference demand expands, operators are trying to add capacity in smaller increments and in more locations. Modular designs promise to standardize power distribution, cooling, networking and compute racks, but they still depend on local utilities, permitting and grid connections. The factory-built model can compress the equipment and integration phases; it cannot remove the external constraints that have delayed projects across Europe and the United States.

For European customers, the appeal may be especially strong. The region has ambitious AI and sovereign compute goals, but it also faces rising electricity costs, grid congestion and lengthy approval processes. A modular product that can be replicated across sites could help operators plan capacity more predictably. It could also make smaller deployments more feasible, allowing universities, enterprises and regional providers to add AI compute without committing to a large, bespoke construction program. The DSX POD concept therefore fits a wider industry pattern: AI infrastructure is moving toward standardized building blocks, even as the underlying chips and workloads remain highly specialized.

Power and water efficiency become board-level metrics

Nvidia's infrastructure message was not only about speed. The company also focused on the resource intensity of AI data centers, which have become a political and operational issue as power demand rises. Its 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% less power. That claim is significant because power availability, not chip supply, has become a bottleneck for many planned AI campuses.

Liquid cooling is increasingly important for high-density AI racks. Air cooling struggles to remove heat from systems packed with power-hungry accelerators, and liquid cooling can transfer heat more efficiently. Closed-loop systems are designed to reduce water loss and contamination risks. Dry coolers, meanwhile, reduce reliance on evaporative cooling, which consumes water. 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 figure illustrates the scale of modern AI campuses: a single large deployment can consume water at rates comparable to small municipalities.

The efficiency debate is not purely environmental. Data center operators face local opposition, utility interconnection queues and corporate sustainability targets. In Europe, new facilities must navigate national energy policies, water-use regulations and community concerns. Technologies that reduce power per token or water per megawatt can make projects easier to approve and cheaper to operate. They can also affect where AI facilities are built. Regions with constrained water supplies may favor dry cooling, while regions with high electricity prices may prioritize power-management software that shifts or throttles workloads.

Still, efficiency gains can be offset by growth. If each token costs less power but total AI usage rises rapidly, aggregate electricity demand may continue to climb. That dynamic has made efficiency a moving target. Nvidia's claims are therefore best understood as a way to improve the economics and siting flexibility of AI infrastructure, not as a solution to the sector's overall resource footprint. The companies buying these systems will need to measure performance across utilization, cooling efficiency and workload management rather than relying on a single benchmark.

OpenShell pushes security into autonomous agent operations

As AI systems move from answering questions to executing tasks, security controls become more complex. 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 architecture builds on Nvidia OpenShell secure runtime software and extends it into full-stack, industrial-scale infrastructure for agent development and operations. The company says the system is designed to enforce security policies as agents execute tasks across enterprise infrastructure and edge systems. The reference architecture is available now.

The timing reflects a wider transition in enterprise AI. Early deployments focused on chatbots and copilots that generated text or recommendations. Agentic systems are different: they can call tools, access databases, trigger workflows and interact with other software. That capability creates value, but it also expands the attack surface. An agent with permission to modify records, send messages or provision resources must be governed by identity, policy and monitoring controls. Security teams need to know what an agent is allowed to do, what it actually did and whether its actions can be reversed.

OpenShell's combination of compute, CPUs, DPUs and software points to an infrastructure-level approach. DPUs can offload networking, storage and security functions from main processors, creating isolation between workloads and control planes. Vera Rubin compute provides the AI processing foundation, while Vera CPUs handle general-purpose tasks. The security software layer is intended to enforce policies across these components. For enterprises, the appeal is a reference design that can be deployed across data centers and edge locations rather than assembled piece by piece.

The agent security market is still forming. Standards for agent identity, permissioning and audit trails are immature, and many organizations are experimenting with frameworks from multiple vendors. A reference architecture can accelerate adoption by giving infrastructure teams a validated blueprint. It can also lock customers into a particular technology stack. The competitive battle will therefore hinge on interoperability, policy flexibility and the ability to monitor agents across hybrid environments. Nvidia's announcement signals that agent security is moving from a software-only concern to a systems design problem.

Robotics platform 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. Simulation is central to that effort because training robots in physical environments is slow and expensive. Synthetic data, physics-based simulation and learning frameworks can help developers test behaviors before hardware is deployed.

Germany's Agile Robots introduced Agile ONE SN, a humanoid reference platform built on Nvidia Isaac GR00T technology and powered by Jetson Thor. The system has 128GB of unified memory and is scheduled to become available in early 2027. The specification is notable because humanoid robots require substantial onboard compute for perception, planning and control. Unified memory can simplify data movement between CPU and GPU workloads, which matters for real-time robotics. The early 2027 timeline also suggests that humanoid platforms are moving from research prototypes toward commercial reference designs, even if broad deployment remains years away.

The robotics opportunity is often framed around manufacturing and logistics, but the technical challenges are broad. Robots must operate safely around people, adapt to unstructured environments and learn tasks without extensive reprogramming. Simulation helps, but the sim-to-real gap remains a hurdle. Developers need accurate sensors, robust models and runtime systems that can respond within strict latency limits. Nvidia's strategy is to provide the computing platform, simulation tools and learning frameworks that let robotics companies focus on their specific applications. Agile Robots' humanoid reference platform is an example of that division of labor: Nvidia supplies the AI and computing foundation, while the robot maker integrates hardware and application expertise.

Telecom operators test AI-powered networks

Nvidia and Nokia announced expanded work with telecom operators on AI radio access networks, or AI-RAN. 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. The announcements point to a longer-term shift in mobile networks, where AI workloads may run alongside traditional radio processing.

Radio access networks are the part of a mobile network that connects devices to the core. They involve complex signal processing, spectrum management and real-time scheduling. AI-RAN explores whether AI accelerators can improve spectral efficiency, reduce energy use or enable new services at the edge. Telecom operators are attracted by the possibility of using the same infrastructure for AI inference and network functions, but they are also cautious about reliability, latency and integration with existing equipment. A commercial call is an early milestone, not a full-scale rollout.

The pilots with BT, Deutsche Telekom and Elisa will test how Nokia software accelerated by Nvidia performs in live operator environments. Each company faces different spectrum holdings, network architectures and regulatory conditions, so the trials may produce varied results. The broader question is whether AI-RAN can deliver measurable gains without increasing operational complexity. Telecom networks are among the most demanding real-time systems in the world, and AI must meet strict availability and safety requirements. The expanded collaboration shows that vendors and operators are willing to move beyond lab trials, but commercial deployment will depend on proven economics and performance. The participating operators are expected to evaluate integration, efficiency and service quality as the pilots progress.


Source: Network World News


Share:

Your experience on this site will be improved by allowing cookies Cookie Policy