NVIDIA’s GTC Berlin announcements cover containerized Vera Rubin infrastructure, a 2.5x substrate supply expansion, plus software for AI factory power management and agent security, alongside new telecom, autonomous vehicle, robotics and healthcare initiatives. Its flagship reveal is the DSX POD reference design, which cuts AI factory deployment on a prepared site from six months to as little as 14 days; partner-built systems will ship in early 2027.

DSX POD: Factory-Integrated AI Infrastructure

DSX POD packages Vera Rubin compute, networking, cabling, rack power and liquid cooling into prefabricated containers. Partners complete all rack wiring, piping and validation before shipment, so delivered units only hook up to site power, cooling and networks. The fast timeline stems from off-site integration and testing, and excludes construction of the supporting facility. Global manufacturing partners include Anduril, Armada, Bull, Crusoe, DayOne, Dell, Delta, Eaton, Foxconn Ingrasys, GIGABYTE, HPE, Lenovo, Pegatron, QCT, Schneider Electric, Vertiv and Wistron. Dell’s matching modular DSX POD rollout shares the same early 2027 schedule. NVIDIA has not published POD capacity, rack counts or pricing.

Expanded Packaging, Substrate, Test and Optical Engine Capacity

NVIDIA will scale substrate output by 2.5x across three years. A multibillion-dollar reservation with IBIDEN triples its Japanese Ono plant, with extra volume ramping in Q1 2028. Samsung Electro-Mechanics joins the substrate supply chain, adding output from existing Korean sites plus new upgrades and factories in Sejong and Busan. Unimicron’s Taiwan KF substrate expansion, backed by a reservation deal, starts volume in Q2 2027 and hits full capacity mid‑2028. Jensen Huang noted the industry is building the largest infrastructure footprint in history.

NVIDIA and Amkor will raise South Korean interposer wafer capacity over 300% by late 2028. Total semiconductor test capacity doubles between 2026 and 2027 via U.S., Taiwan and Singapore expansions. A $1.1 billion reservation with KYEC funds new U.S. test facilities targeting high-volume output in April 2027, while expanded Singapore test lines ramp in February 2027. For co-packaged optics, NVIDIA and SPIL plan a 15x jump in Taiwan optical-engine manufacturing by mid‑2027, followed by another doubling by the end of 2028. All figures are forward-looking targets, with no baseline volumes disclosed.

DSX Full-Stack Architecture: Power, Cooling and Cluster Tooling

DSX is NVIDIA’s full-stack reference architecture unifying compute, networking, storage, software, power and cooling. Its software suite includes DSX MaxLPS for dynamic power allocation, DSX Flex for grid-aware workload scheduling, DSX Sim for simulation, DSX OS for infrastructure management, and DSX Exchange for facility-system integration.

Vera Rubin supports full liquid cooling at coolant temperatures up to 45°C, paired with closed-loop dry-cooler heat rejection that can reduce site cooling water use nearly to zero, with only supplemental mechanical cooling needed in extreme heat. At 2 GW scale, NVIDIA estimates daily water savings of 54 million liters versus traditional cooling towers.

DSX MaxLPS can sustain equal token throughput with up to 30% less power. Lambda testing on HGX B200 delivered 24% higher token throughput and 19% more usable GPU capacity under a fixed power cap. An Nscale Iceland trial across four racks recorded 49% higher inference throughput and 37% more managed GPU capacity within a 264 kW budget. These are vendor-provided results not independently audited.

DSX Flex aligns workloads with utility pricing, demand response, on-site generation and battery storage. A joint trial with Emerald AI, EPRI, Nebius, UK National Grid and NVIDIA reduced a Blackwell Ultra cluster’s power draw by up to 40% during simulated grid stress while preserving critical jobs.

The new SimReady Central Library supplies validated power and cooling assets for DSX digital twins. Nscale uses DSX Sim in Portugal and Iceland, while Schneider Electric, AVEVA, Siemens, Dassault Systèmes, Vertiv, ABB, PTC and Procore contribute equipment models, simulation and construction workflows; ABB leverages its Infinitus DC portfolio for 800 VDC development.

Three DSX OS modules enter developer preview: Config Manager for networking, Rack Management Service for rack power, inventory, firmware and switch control, and Cluster Readiness Engine for distributed validation and node fault isolation. They complement NVIDIA Infrastructure Controller, AI Cluster Runtime and NVSentinel. Trane released seven thermal reference designs including three 250 MW DSX cooling layouts. Vertiv and LiquidStack coolant distribution units, plus Hitachi Energy battery storage, have earned DSX Ready qualification.

AI-RAN Extends Accelerated Infrastructure to Telecom

NVIDIA and Nokia are commercializing shared platforms that run RAN software and AI workloads on the same accelerated hardware. Roughly 20 operators have announced AI-RAN partnerships built on Nokia anyRAN and NVIDIA Aerial ARC servers. T-Mobile completed what NVIDIA calls the world’s first commercial AI-RAN 5G Standalone call in Bellevue using ARC, CUDA-accelerated Nokia RAN stack, massive MIMO radios and standard handsets. Elisa demonstrated concurrent voice, data and video analytics over the air, clearing a path for 2027 commercial AI-RAN services. BT will run Birmingham trials in 2027; Deutsche Telekom validates in Bonn and Oulu ahead of field tests; Orange tests the stack for spectral efficiency and edge services.

Two new ARC variants, MGX 2U and MGX 4U, use RTX PRO 6000 Blackwell Server GPUs with x86 or Vera CPU options, sold by Dell, QCT and Supermicro; no ship date is given. Deutsche Telekom expands its Industrial AI Cloud for distributed robotics. Orange deploys ARC for Nemotron speech services, and Elisa has an ARC enterprise voice assistant. New deployments span CRA, Etisalat, Silknet and TIM, with expansions for Cassava, Fastweb+Vodafone, Kazakhtelecom and Orion Telecom. Nokia embeds Nemotron and the NVIDIA Agent Toolkit into its network stack; Cisco integrates tuned Nemotron models into Crosswork AI for optical and IP transport; TCS, BT and NTT DATA build agents for inspection, fault remediation and early degradation detection.

OpenShell: Hardware-Enforced Agent Safety

The OpenShell reference architecture is now available. It pairs Vera Rubin for model execution, Vera CPUs hosting agent harnesses and the OpenShell runtime, plus BlueField DPUs running NVIDIA Sentry for out-of-band monitoring and silicon-level policy enforcement. Jensen Huang stated AI capability and safety must advance in tandem.

Built on DOCA for BlueField-4, Sentry isolates and halts misbehaving agents within milliseconds when they cross access boundaries. NVIDIA claims up to 20% higher agent density with Sentry, 30x higher per-megawatt throughput for Vera Rubin NVL72 agent workloads, and double Vera CPU performance versus x86, without published test conditions. The stack runs on Linux distributions from Canonical, SUSE and Red Hat, Windows via Microsoft Execution Containers, and major public clouds, with edge support on RTX Spark, DGX Spark and DGX Station. SAP integrates OpenShell into Joule Studio. ServiceNow Arc launched generally available, using OpenShell and Sentry for runtime governance, with its AI Passport able to revoke agent privileges across hardware and software in seconds. Dell, Cisco, HPE, Lenovo and Supermicro build certified hardware. CoreWeave, Crusoe, Firebird, GMI Cloud, IREN, Lambda, Nebius, Nscale, OCI and Vultr develop cloud offerings. Security vendors including 1Password, Check Point, CrowdStrike, F5, Fortinet, Microsoft, Okta, Palo Alto Networks, TrendAI and Zscaler add OpenShell integration.

Hyperion Autonomous Vehicle Updates

Volvo adopts NVIDIA Hyperion and the Alpamayo Stack; Mercedes extends the platform toward L3/L4; Volkswagen Commercial Vehicles showcases ID. Buzz AD on Hyperion; Einride builds its autonomous freight stack using Halos, Alpamayo and Cosmos. Bolt selects Alpamayo for a European robotaxi fleet based on future Hyperion-powered Lucid vehicles, targeting Madrid data collection in 2027 pending regulation and a rollout of at least 25,000 L4 vehicles.

NVIDIA launched Halos SAFE, a safety assurance framework linking simulation, hardware-in-the-loop, road testing and fleet monitoring into auditable safety evidence, with dSPACE, IVEX and TÜV Rheinland as partners. Alpamayo 2 Edge is a 2B-parameter reasoning model built on Cosmos 3 Edge, tuned for DRIVE AGX Thor and 10 Hz in-vehicle perception and planning. New tooling includes AlpaJudge for driving explanation scoring, Foundation4D for lidar-free multi-camera labeling, and compact Cosmos 3 Nano synthetic generation models.

Cosmos Data Factory converts raw driving logs into labeled datasets. Alpamayo Agentic Factory, arriving early 2027, unifies data prep, post-training and simulation evaluation. NIM containers now support Alpamayo 1.5 and NuRec Harmonizer. Porsche leads a Volkswagen Group Azure workflow combining Voxel51, Cosmos Data Factory and Omniverse NuRec, projecting up to 90% fewer physical test drives and an 85% shorter development cycle.

Isaac Robotics: Humanoid Reference and New Runtime

Agile Robots’ Agile ONE SN reference humanoid, powered by Jetson Thor, is scheduled for early 2027. The 1.4 m, 40 kg Agile ONE S has 31 degrees of freedom, 3 kg per-arm payload, dual five-finger WUJI Hand 2 grippers with 1 kHz force control and tactile feedback. Its onboard compute delivers 2,070 FP4 TFLOPS and 128 GB unified memory, with multi-modal sensing; NVIDIA will tie its safety controls into Halos.

Isaac Lab 3.0 is generally available open source with multi-GPU pipelines and Newton/PhysX physics engines. Isaac OS, a Holoscan-based open runtime, delivers deterministic real-time control linking AI, sensors and robot actuators. The SimReady Central Library hosts over 2,000 factory and robotics assets, and the Human 1M Dataset offers synchronized multi-view video and 3D hand tracking, both published on Hugging Face; full dataset access is reserved for Cosmos Coalition members.

Three open DGX Station workflows support GR00T 1.7 reinforcement learning, dexterous manipulation training and Cosmos synthetic data post-training. Halos on IGX Thor provides the safety foundation for industrial and warehouse robots. European adoption includes ABB integrating Omniverse with RobotStudio HyperReality, KUKA using it in Visual Components FactoryLens, Sereact training handling models on GR00T plus Cosmos Transfer, and Toyota Material Handling Europe and SoftServe cutting deployment time by up to 50% with Omniverse and Isaac Sim.

BioNeMo and Healthcare Expansion

NVIDIA and UW Medicine’s Institute for Protein Design released RFD4-Proteina, an open BioNeMo model for programmable protein design, searching and refining candidates from molecular constraints and scoring targets. Benchmarks show improved design success and throughput across binding, enzyme scaffolding and DNA design tasks, though lab validation remains required. The model is available via Amazon Bio Discovery, Anthropic Claude Science, Benchling, Databricks, Google Cloud, Hugging Face, Nebius, OpenAI and Tamarind Bio.

Siemens Healthineers and NVIDIA build adaptive radiotherapy workflows combining agentic co-planning, patient simulation and real-time compute, using Nemotron for planning, CUDA dose calculation and Holoscan for live sensor fusion. Benchling, Biolevate, Boltz, Automata, Genedata, Hologen and Therna Biosciences develop scientific agents and lab workflows on BioNeMo and Nemotron; Genedata’s Danaher autonomous lab foundry will embed RFD4-Proteina with experimental feedback. Avelios, Corti, Dedalus, Dawn Health and Sword build Nemotron-powered clinical scribes and care assistants. PolyAI deploys multilingual medical dialogue with Corilus for over 70,000 clinicians in France and Belgium. Apian and VitVio construct hospital digital twins using Cosmos and Omniverse.

Beijing Qianxing Jietong Technology Co., Ltd.
Sandy Yang/Global Strategy Director
WhatsApp / WeChat: +86 13426366826
Email: yangyd@qianxingdata.com
Website: www.qianxingdata.com/www.storagesserver.com
Business Focus:
ICT Product Distribution/System Integration & Services/Infrastructure Solutions
With 20+ years of IT distribution experience, we partner with leading global brands to deliver reliable products and professional services.
“Using Technology to Build an Intelligent World”Your Trusted ICT Product Service Provider!