NVIDIA’s Berlin Plans Tie AI Factories to Working Robots

NVIDIA wants AI everywhere. At GTC Berlin 2026, the company outlined plans spanning infrastructure, software, semiconductor manufacturing, telecom networks, enterprise agents, robotics and autonomous vehicles. The pitch is not just more computing hardware—it is a connected buildout for systems that produce tokens, move data and operate in the physical world.

“We are in the midst of the largest infrastructure buildout in human history. Demand is extraordinary. Capacity is the constraint,” said Jensen Huang, NVIDIA’s Founder and CEO. The company announced agreements covering substrates, advanced packaging, semiconductor testing and optical connectivity, with plans to double total semiconductor testing capacity between 2026 and 2027.

NVIDIA plans to increase substrate capacity by 2.5 times over the next three years, while its agreement with Amkor targets more than 300% growth in advanced packaging capacity by the end of 2028. The company also presented DSX POD, which could reduce deployment time from six months to as little as 14 days, and DSX MaxLPS, which aims to deliver the same AI output using up to 30% less power.

Cooling also joined the capacity conversation. NVIDIA estimates that its cooling design could save 54 million litres of water daily at a 2 GW deployment—a number large enough to make “sustainable infrastructure” sound less like a conference badge and more like an engineering requirement.

Telecom networks become AI infrastructure

NVIDIA announced that T-Mobile completed the world’s first AI-RAN call on a commercial network. Finnish telecom player Elisa completed Europe’s first field trial of AI-RAN, running voice, mobile data and vision AI on one platform over 5G.

Nine telecom operators across Europe, the Middle East and Africa are building or expanding NVIDIA-powered infrastructure. BT and Deutsche Telekom are among the telecom operators named in the broader deployment effort, as networks move beyond carrying communications alone.

“Telecom networks were built to carry the world’s conversations. Now, they will also carry the world’s intelligence,” Huang said. That shift puts AI workloads inside network infrastructure, rather than treating the network as a passive connection between users and distant computing systems.

NVIDIA also expanded its open-source Agent Toolkit with the Agent Knowledge Foundation, Nemotron Retriever models and the Kumo family. Those additions target enterprise agents that can use knowledge and retrieval tools, extending NVIDIA’s focus from AI infrastructure into software that performs work.

A reference humanoid with a rather specific specification sheet

Agile Robots announced Agile ONE SN, a humanoid reference platform built on NVIDIA Isaac GR00T and powered by NVIDIA Jetson Thor. NVIDIA says it will be available in early 2027, giving developers a full system for creating, training and validating humanoid capabilities before the hardware arrives in the workplace.

“Today, Agile Robots introduced Agile ONE SN, a full-stack, open reference humanoid built on NVIDIA Isaac GR00T, powered by Jetson Thor, giving developers a ready system to create, train and validate the capabilities needed to put humanoids to work,” said Zhaopeng Chen.

Agile ONE SN stands 1.40 meters tall, weighs 40 kilograms, has 31 degrees of freedom and supports a 3-kilogram payload per arm. It includes dual WUJI Hand 2 five-finger hands with 20 degrees of freedom, 1kHz force control and tactile sensing, alongside multimodal sensing and planned integration of NVIDIA Halos safety capabilities.

Jetson Thor provides 2,070 FP4 teraflops of compute and 128GB of unified memory. The platform also includes onboard compute, the Isaac GR00T platform and a humanoid body designed around NVIDIA’s broader robotics stack—not a research demo balanced on a carefully edited video.

Chen described the goal plainly: “The breakthrough for humanoids is usefulness at scale: one adaptable machine that learns many skills and works across environments built for people.” The reference platform is one part of a major Isaac expansion that includes a reference humanoid, datasets, runtime updates and changes to NVIDIA’s robot learning framework.

NVIDIA Isaac now adds the Isaac OS runtime, Isaac Lab 3.0 and open-source reference workflows. Isaac OS connects AI, sensors, controls and robot hardware for real-time performance, while Isaac Lab 3.0, now generally available, provides an open-source, agent-ready framework for training and refining robot policies.

NVIDIA’s open-source robot learning reference workflows run on DGX Station and have support from various partners. NVIDIA Halos adds a safety platform built on NVIDIA IGX Thor, giving the robotics stack a defined place for safety alongside perception, control and learning.

European robotics companies including ABB Robotics, KUKA, Sereact, Humanoid, Hexagon Robotics and Bear Robotics are building on NVIDIA’s open technologies. Toyota Material Handling Europe and SoftServe use Omniverse and NVIDIA Isaac Sim for autonomous material-handling solutions, reducing deployment time by up to 50%.

Other partnerships connect the platform to specific robot tasks. Hugging Face and Pollen Robotics use Isaac Lab, Isaac Sim and Newton to build Microduck; Franka Robotics uses NVIDIA NIM microservices and Isaac Sim to turn natural-language requests into actions; and Bellboy uses Jetson and GR00T to power Billie, a robot deployed at hotels operated by Marriott, Hilton, IHG and OYO.

Schaeffler contributes actuator technology for NVIDIA Isaac Lab and Isaac Sim, while NEURA Robotics uses Isaac Sim and Isaac Lab to generate synthetic training data and train robot skills. NVIDIA’s Berlin announcements point to one strategy: expand the machinery first, then let developers and industrial partners decide where the machines earn their keep.

Stay connected via Google News

Follow us for the latest AI News