# NVIDIA GTC Berlin: 14-Day DSX POD AI Factories, a 2.5x Substrate Buildout, and the OpenShell Reference Architecture > Source: > Published: 2026-10-21 18:04:42+00:00 NVIDIA’s GTC Berlin announcements today span containerized Vera Rubin infrastructure, a 2.5x expansion of its substrate supply, and software for AI factory power management and agent security, along with telecom, autonomous vehicle, robotics, and healthcare programs. The headline number is a deployment target: NVIDIA says its new DSX POD reference design cuts an AI factory’s deployment at a prepared site from six months to as little as 14 days, with partner-built systems due in early 2027. ## DSX POD Moves Infrastructure Integration into the Factory NVIDIA DSX POD combines Vera Rubin compute with networking, cabling, rack-level power, and liquid cooling in factory-built containers. Partners complete rack-to-rack cabling, wiring, piping, and functional validation before shipment, so a delivered container connects to prepared power, cooling, and networking at the site. NVIDIA attributes the shorter window to finishing integration and testing at the manufacturing facility; the 14-day figure assumes a prepared site and leaves out the time to build the supporting facility. Partners targeting worldwide delivery in early 2027 are Anduril, Armada, Bull, Crusoe, DayOne, Dell Technologies, Delta, Eaton, Foxconn (Ingrasys), GIGABYTE, HPE, Lenovo, Pegatron, Quanta Cloud Technology (QCT), Schneider Electric, Vertiv, and Wistron. [Dell’s version](https://www.storagereview.com/news/dell-ai-factory-nvidia-poweredge-m9882-vera-cpu-rack-xe9810-groq-3-gtc-berlin), the Dell Modular Data Center with NVIDIA DSX POD, was announced this morning with the same early 2027 timing. NVIDIA has not published per-POD capacity, rack counts, or pricing. ## Manufacturing Plans Expand Substrates, Packaging, Testing, and Optical Engines NVIDIA plans to increase substrate capacity for its products by 2.5x over the next three years. A multibillion-dollar capacity reservation advance to IBIDEN supports expansion of IBIDEN’s Ono Plant in Japan to three times its current capacity, with additional capacity expected to begin ramping in Q1 2028. Samsung Electro-Mechanics joins NVIDIA’s substrate supply chain with capacity from its existing South Korean operations plus facility upgrades and new factories in Sejong and Busan. NVIDIA is also supporting Unimicron’s KF substrate expansion in Taiwan through a capacity reservation agreement, with capacity arriving from Q2 2027 and full capacity by mid-2028. “We are in the midst of the largest infrastructure buildout in human history,” said Jensen Huang, founder and CEO of NVIDIA. NVIDIA and Amkor plan to increase interposer wafer capacity in South Korea by more than 300 percent by the end of 2028. NVIDIA expects its total semiconductor testing capacity to double from 2026 to 2027 through expansion in the U.S., Taiwan, and Singapore; a $1.1 billion capacity reservation advance to King Yuan Electronics (KYEC) funds new U.S. testing facilities and equipment, with high-volume production targeted for April 2027, and expanded Singapore testing operations target high-volume production in February 2027. For co-packaged optics, NVIDIA and Siliconware Precision Industries (SPIL) plan to expand Taiwanese optical-engine manufacturing capacity by 15x by mid-2027, then double it again by the end of 2028. Every figure in this section is a planned capacity target, and NVIDIA gives no current baseline volumes. ## DSX Addresses Cooling, Power Budgets, and Cluster Readiness NVIDIA DSX is the full-stack reference architecture that combines compute, networking, storage, software, power, and cooling. Its software components are DSX MaxLPS for dynamic power allocation within a provisioned budget, DSX Flex for coordinating workload power with grid conditions, DSX Sim for simulation, DSX OS for infrastructure management, and DSX Exchange for connecting computing and facility systems. Vera Rubin infrastructure supports 100 percent liquid cooling with coolant temperatures up to 45 degrees Celsius, and the DSX reference designs connect it to closed-loop, dry-cooler heat rejection. NVIDIA says those designs can cut facility cooling water consumption to near zero, with local climate deciding whether supplemental mechanical cooling is needed in the hottest periods. At a 2-gigawatt scale, NVIDIA estimates savings of 54 million liters of water per day against conventional cooling-tower systems. NVIDIA says DSX MaxLPS can deliver the same token output using up to 30 percent less power. In a Lambda proof of concept on HGX B200 systems, the company reports 24 percent higher cluster token throughput and 19 percent more GPU capacity within a fixed power budget, and an Nscale evaluation across four [GB300 NVL72](https://www.storagereview.com/news/meta-prometheus-gigawatt-supercluster-1-million-gpus-gb200-gb300-nvl72) racks in Iceland reports 49 percent higher inference throughput and 37 percent more managed GPU capacity within the same 264-kilowatt provisioned budget. These are vendor results, and we did not run or audit them. DSX Flex coordinates workload demand with utility conditions, demand-response events, energy prices, onsite generation, and storage. In an early trial with Emerald AI, EPRI, Nebius, the U.K. National Grid, and NVIDIA, a Blackwell Ultra cluster reduced power draw by up to 40 percent during simulated grid events while critical workloads kept running. NVIDIA’s new SimReady Central Library provides validated power and cooling equipment assets for DSX digital twins. Nscale uses DSX Sim at facilities in Portugal and Iceland, and Schneider Electric, AVEVA, Siemens, Dassault Systèmes, Vertiv, ABB, PTC, and Procore are contributing equipment models, simulation, or construction workflows; ABB’s work also covers [800 VDC](https://www.storagereview.com/news/nvidia-moves-800-vdc-power-architecture-from-concept-to-production-just-dont-turn-off-ac-power-yet) development through its Infinitus Direct Current portfolio. Three DSX OS components enter developer preview: NVIDIA Config Manager for network configuration, NVIDIA Rack Management Service for rack-level power, inventory, firmware, and switch management, and NVIDIA Cluster Readiness Engine for distributed validation and node-level fault isolation. They join NVIDIA Infrastructure Controller, NVIDIA AI Cluster Runtime, and NVSentinel. On the facility side, Trane Technologies has published seven thermal management reference designs, three of them 250-megawatt DSX-based cooling designs. Trane’s LiquidStack GigaModular and Vertiv’s 2.3-megawatt CoolChip coolant distribution units have qualified in their [DSX Ready](https://www.storagereview.com/news/nvidia-dsx-ready-qualifies-tesla-lg-and-hitachi-batteries-and-vertiv-liquidstack-and-lg-cdus-for-ai-factories) categories, and Hitachi Energy’s battery energy storage system has earned DSX Ready BESS qualification. ## AI-RAN Extends GPU Infrastructure into Telecom Networks NVIDIA and Nokia are moving platforms that run radio access network software and AI services on shared accelerated infrastructure toward commercial service. NVIDIA says nearly 20 operators have publicly announced AI-RAN collaborations using Nokia anyRAN software and NVIDIA Aerial RAN Computer (ARC). T-Mobile completed what NVIDIA describes as the world’s first AI-RAN call on a commercial network, in Bellevue, Washington, using commercial ARC servers, Nokia’s CUDA-accelerated RAN software, massive MIMO radios, an existing 5G Standalone core, and a commercial smartphone. Elisa ran voice, mobile data, and video analytics at the same time on ARC in an over-the-air 5G field environment, which NVIDIA says clears the way for commercial AI-RAN services in 2027. BT plans live trials in Birmingham in 2027, Deutsche Telekom plans field trials after validating the platform in Bonn and Oulu, and Orange is testing Nokia 5G software on ARC for spectral efficiency and edge services. Two new ARC configurations, MGX 2U and MGX 4U, use NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs and will offer x86 or NVIDIA Vera CPU options. Dell Technologies, QCT, and Supermicro will offer ARC-ready servers; NVIDIA gives no ship date for either configuration. Deutsche Telekom is expanding its Industrial AI Cloud and demonstrating robotics workloads distributed across onboard, edge, and central infrastructure. Orange is deploying ARC for voice applications on Nemotron Speech, and Elisa has deployed an ARC-based enterprise voice assistant. NVIDIA also lists new infrastructure projects at CRA, Etisalat, Silknet, and TIM, and expansions at Cassava, Deutsche Telekom, Fastweb + Vodafone, Kazakhtelecom, and Orion Telecom. Nokia is integrating Nemotron and the NVIDIA Agent Toolkit into its [Autonomous Networks Suite](https://www.storagereview.com/news/nvidia-pushes-telco-toward-autonomous-networks-with-open-nemotron-ltm-and-new-blueprints), Cisco has put fine-tuned Nemotron models into Crosswork AI for IP and optical transport operations, and TCS, BT, and NTT DATA are building agents for installation inspection, fault diagnosis, remediation, and early detection of network degradation. ## OpenShell Separates Agent Execution from Hardware Security Enforcement The [NVIDIA OpenShell](https://www.storagereview.com/news/nvidia-open-agent-safety-platform-openshell-sentry-bluefield-4) Reference Architecture is available now. It pairs Vera Rubin infrastructure for model workloads with Vera CPUs running agent harnesses and the OpenShell runtime, and BlueField DPUs running NVIDIA Sentry for out-of-band monitoring and in-silicon policy enforcement. “AI capability and safety must advance together,” said Jensen Huang, founder and CEO of NVIDIA. NVIDIA says Sentry, built on DOCA and running on BlueField-4, quarantines and stops an agent within milliseconds when it tries to move outside its software boundary. The company also claims up to 20 percent higher agent capacity with Sentry, 30x higher throughput per megawatt for agentic workloads on Vera Rubin NVL72, and twice the performance of x86 processors for Vera CPUs, with no comparison configurations or workload conditions given for any of the three. The architecture runs on Linux and Kubernetes from Canonical, SUSE, and Red Hat, on Windows through Microsoft Execution Containers (MXC), and on AWS, Google Cloud, and Microsoft Azure, with edge deployments on RTX Spark, DGX Spark, and DGX Station. SAP is integrating OpenShell with the Joule Studio runtime. ServiceNow announced general availability of ServiceNow Arc, with runtime governance through its Super Intelligence Control Tower and enforcement through OpenShell; its AI Passport authority layer extends policies to OpenShell and Sentry and can revoke an agent’s access across software and hardware within seconds. Dell Technologies, Cisco, HPE, Lenovo, and Supermicro are building infrastructure on the architecture. CoreWeave, Crusoe, Firebird, Firmus, GMI Cloud, IREN, Lambda, Nebius, Nscale, Oracle Cloud Infrastructure, and Vultr are working on cloud deployments, and 1Password, Check Point, CrowdStrike, Cisco, F5, Fortinet, Microsoft, Okta, Palo Alto Networks, TrendAI, and Zscaler are integrating support into their agent security products. ## Hyperion Expands Vehicle Programs and Safety Validation Volvo Cars is adopting NVIDIA Hyperion and the Alpamayo Stack; Mercedes-Benz is extending its work on the platform to L3 and L4 capabilities; Volkswagen Commercial Vehicles is showing its ID. Buzz AD platform on Hyperion, and Einride is building its autonomous freight driver on Hyperion with Halos, Alpamayo, and Cosmos. Bolt has selected the Alpamayo Stack for a European robotaxi program that starts with future Hyperion-based Lucid vehicles, with driving data collection in Madrid in 2027 subject to regulatory approval and a planned rollout of at least 25,000 L4-capable Lucid vehicles. NVIDIA introduced Halos SAFE, the Safety Assurance Framework for Evaluation, to connect simulation, software- and hardware-in-the-loop evaluation, road testing, and fleet monitoring into traceable safety evidence, with dSPACE, IVEX, and TÜV Rheinland as collaborators. Alpamayo 2 Edge is a 2-billion-parameter reasoning model built on Cosmos 3 Edge’s Nemotron-based foundation, optimized for DRIVE AGX Thor and designed to run at 10Hz for in-vehicle scene understanding and motion planning. AlpaJudge scores driving explanations, Foundation4D labels multicamera video without lidar input, and new Cosmos 3 Nano variants add synchronized multiview, multicamera, and lidar generation. Cosmos Data Factory’s AV blueprint turns driving video and vehicle signals into curated, labeled datasets, and Alpamayo Agentic Factory, due in early 2027, will connect data preparation, post-training, and simulation evaluation. NIM inference containers now cover Alpamayo 1.5 and NuRec Harmonizer. Porsche is leading a Volkswagen Group data workflow on Microsoft Azure through Voxel51, Cosmos Data Factory, and Omniverse NuRec, which Porsche projects will cut repeated physical test drives by up to 90 percent and development-cycle duration by up to 85 percent. ## Isaac Adds a Reference Humanoid and Robot Runtime Agile Robots’ Agile ONE SN reference humanoid pairs [Isaac GR00T](https://www.storagereview.com/news/nvidia-computex-2026-keynote-the-rtx-spark-pc-family-dgx-station-and-physical-ai) with Jetson Thor and is scheduled for early 2027. The Agile ONE S body stands 1.40 meters, weighs 40 kilograms, and provides 31 degrees of freedom with a 3-kilogram payload per arm; its dual WUJI Hand 2 five-finger hands each offer 20 degrees of freedom, 1kHz force control, and tactile sensing. Onboard [Jetson Thor](https://www.storagereview.com/news/nvidia-launches-jetson-agx-thor-developer-kit) compute delivers a stated 2,070 FP4 teraflops and 128GB of unified memory; sensing comes from head and wrist cameras, lidar, and inertial measurement units; and NVIDIA plans to integrate on-robot safety protocols with Halos. Isaac Lab 3.0 is generally available as an open-source learning framework with multi-GPU workflows and the Newton and PhysX physics engines, and NVIDIA introduced Isaac OS, an open-source runtime built on Holoscan that connects AI, sensors, controls, and robot hardware for deterministic real-time operation. SimReady Central Library carries more than 2,000 robotics and factory assets, and the Human 1M Dataset holds synchronized three-view video and 3D hand tracking; both are on Hugging Face, with the full Human 1M Dataset open to Cosmos Coalition partners. Three open-source DGX Station reference workflows cover reinforcement-learning refinement of GR00T 1.7, dexterous manipulation training in Isaac Lab, and Cosmos post-training for synthetic data generation using ChatGPT Astra. Halos, built on IGX Thor, is the safety platform alongside them, with robotics partners integrating or evaluating it for industrial and warehouse use. In Europe, ABB Robotics is integrating Omniverse libraries with RobotStudio HyperReality, KUKA is using them in Visual Components FactoryLens, Sereact is training its goods-handling platform on GR00T with Cosmos Transfer, and Toyota Material Handling Europe and SoftServe report deployment-time reductions of up to 50 percent with Omniverse and Isaac Sim. ## BioNeMo Protein Design and Healthcare Workflows Expand NVIDIA and the UW Medicine Institute for Protein Design released RFD4-Proteina, an open BioNeMo model that combines programmable protein design with search, generating and refining candidates from molecular inputs, constraints, and scoring objectives. NVIDIA says the model’s research paper reports higher computational design success rates and generation throughput than the evaluated baselines across protein binding, small-molecule binding, enzyme scaffolding, and DNA binding benchmarks, without giving the magnitudes, and laboratory testing still decides how any design performs. The model is available today through Amazon Bio Discovery, Anthropic’s Claude Science, Benchling, Databricks, Google Cloud, Hugging Face, Nebius, OpenAI, and Tamarind Bio. Siemens Healthineers and NVIDIA are developing adaptive radiotherapy workflows that combine agentic co-planning, patient-specific simulation, and real-time computing; the collaboration is exploring Nemotron for planning assistance, CUDA-accelerated dose calculation, and Holoscan for connecting live sensor data with AI and simulation. Benchling, Biolevate, Boltz, Automata, Genedata, Hologen, and Therna Biosciences are developing scientific agents, models, or laboratory workflows on BioNeMo, Nemotron, and related tools, and Genedata’s planned protein design foundry inside Danaher’s autonomous lab will incorporate RFD4-Proteina with experimental feedback. Avelios, Corti, Dedalus, Dawn Health, and Sword are building on Nemotron for AI scribes, clinical care assistance, and patient support; PolyAI is deploying multilingual dialog with medical software provider Corilus across more than 70,000 healthcare professionals in Belgium and France; and Apian and VitVio are building hospital digital twins with Cosmos and Omniverse libraries.