I summarized Jensen Huang's keynote at GTC Berlin 2026
This page has been translated by machine translation. View original
Introduction
Hello, I'm Morishige from Classmethod's Manufacturing Business Technology Department.
On October 21, 2026, NVIDIA GTC Berlin held in Berlin, Germany featured a keynote address by CEO Jensen Huang.
Announcements were published one after another on NVIDIA's blog and newsroom throughout the keynote, with the press kit featuring 8 press releases and 9 blog posts. As someone who has been working with DGX Spark and OpenShell, there was a lot to take interest in, and I decided to organize everything including changes since GTC Taipei in June.
This article introduces the main announcements following the flow of the keynote.
The Flow of the Keynote and Changes Since Taipei in June
The keynote began with an opening video depicting tokens as "the product of a new factory that generates intelligence."
The overall flow was as follows.
| Topic | Main Announcements |
|---|---|
| 33 years with Europe and the 5-layer AI cake | Commitment of approximately $10 billion to Europe |
| Full stack and Vera Rubin hardware | Fully liquid-cooled Vera Rubin |
| DSX and DSX POD | AI factory assembled and delivered |
| 5 elements of agents and OpenShell | OpenShell Reference Architecture |
| Open models and Agent Knowledge Foundation | Nemotron 3.5, Kumo |
| Physical AI | Hyperion, Agile ONE SN |
| RTX Spark and DGX Station | PCs for developing agents locally |
| Microduck demo | RTX Spark orchestrates robot training |
| Closing video | Recap of the keynote in song |
Comparing with the keynote at GTC Taipei in June, there are the following differences.
| Theme | Taipei (June) | Berlin (October) |
|---|---|---|
| Agent elements | Model, Harness, Tools, Skills, Runtime | model, harness, sandbox, knowledge, tools |
| OpenShell | Agent runtime | Sandbox within the 5 elements. Reference Architecture published |
| Knowledge | Not in element names | Agent Knowledge Foundation |
| AI factory | Three tiers: RTX, DGX, DSX | DSX POD assembled and delivered from the factory |
| RTX Spark | Announced | Already on sale as of October 16 |
| Regional narrative | Taiwan's supply chain | Europe can become a producer of AI |
The biggest change is the reorganization of element names used to describe agents. Sandbox and knowledge became named elements, and OpenShell was repositioned as part of the sandbox element.
Topics such as full mass production of Vera Rubin and positioning CUDA-X as a tool for agents remain unchanged from Taipei.
A Call to Europe: From Consuming AI to Producing It
The first half of the keynote began with NVIDIA's 33-year relationship with Europe. The history began with creating early chips with Europe's SGS-Thomson and with Germany's ELSA serving as an entry point into the workstation market.
Building on that, AI was organized as a 5-layer cake of "energy, computing, AI factories, models, and applications." NVIDIA has committed approximately $10 billion to Europe's technology ecosystem, with partners exceeding 12,000.
The strongest message of this section was the following statement:
To be part of this AI economy, you don't want to just be consuming the AI; you want to be producing the AI. We could do all of that here in Europe, and in fact, we've already started.
The call was that rather than simply using AI, Europe should become a producer of it, and that this has already begun. Where Taipei had put Taiwan's supply chain front and center, Berlin felt like a keynote that asked which of the 5 layers you want to own for yourself.
Vera Rubin Hardware and the DSX POD Assembled and Shipped from the Factory
On the hardware side, the actual Vera Rubin unit that has entered full mass production was introduced on stage. It's a 100% liquid-cooled tray with no cables or wiring, reportedly reducing assembly effort significantly. CEO Huang emphasized that through extreme co-design—redesigning everything from chips to algorithms simultaneously—AI performance has been multiplied by one million over the past 10 years.
The first new announcement was the NVIDIA DSX POD. It is a reference design in which partners assemble a complete Vera Rubin AI factory inside a modular container, verify its operation, and then deliver it on-site.
According to the press release, at a prepared site, the deployment period can be reduced from 6 months to as little as 14 days. Delivery is planned for early 2027. A considerable portion of the keynote was dedicated to how to shorten the time from delivery to the computer becoming operational.
In addition, it was announced that European partners would adopt the DSX AI factory platform, which can reduce cooling water usage by up to 100% and deliver the same output with up to 30% less power.
Agents Rebuilt Around 5 Elements
In the middle of the keynote, CEO Huang explained agents using 5 elements.
There are five things. There's the model. There's the harness, the sandbox that the harness sits in, and two other things: knowledge and tools.
The 5 elements and their corresponding components are shown in the table below.
| Element | Description in keynote | Corresponding component |
|---|---|---|
| model | Handles perception, reasoning, and planning | Open models like Nemotron, and closed models |
| harness | Exoskeleton of the agent | OpenClaw, Hermes, OpenHands, etc. Custom builds also possible |
| sandbox | Isolates the harness from its surroundings | OpenShell and NVIDIA Sentry |
| knowledge | Connects to internal documents and data | Agent Knowledge Foundation |
| tools | Handles specialized computation | CUDA-X libraries |
The harness was described as "the exoskeleton of the digital agent," with names like OpenClaw, Hermes, and OpenHands mentioned. Hermes Agent, which I've been running on DGX Spark, is one of them.
The OpenShell Reference Architecture press release includes a comment from CEO Huang: "agents can think on Rubin, act on Vera and be governed on BlueField." The division of roles is: inference on Vera Rubin, harness and tool execution on Vera CPU, and governance on BlueField-4.
Constraining Agents from the Outside with OpenShell and Sentry
The sandbox was explained with introductory videos for OpenShell and Sentry.
Please don't delete anything. (...) Please don't touch my SSH keys. That's the please don't approach. But hope isn't security.
This raised the problem that simply asking agents to "not delete things" or "not touch SSH keys" doesn't make them secure.
OpenShell is a runtime that creates a sandbox for each agent and enforces policies at the level of system calls, files, and network access. Keys are held by a supervisor outside the sandbox and are only added to communications with permitted destinations.
NVIDIA Sentry, running on BlueField-4, monitors agent activity through a separate channel. When unintended behavior is detected, it disconnects and isolates the model.
What's newly announced is the NVIDIA OpenShell Reference Architecture, which enables these two components to be integrated across everything from AI factories to RTX Spark and DGX Spark. Partners listed include Anthropic, AWS, Google Cloud, Microsoft, and OpenAI.
There was also a note that permissions cannot be fixed and need to be dynamically changed depending on the task and user. When I tried OpenShell 0.1.0 in September, the only policy that could be changed while running was in the network section. How far dynamic permissions can go is something I want to verify with the Reference Architecture.
(This article is from September 26, 2026. OpenShell at the time was in the 0.1.0 series.)
Connecting Company Data with Open Models and Agent Knowledge Foundation
In the model section, 6 families of open models were introduced: Nemotron, Cosmos, GR00T, Alpamayo, Ising, and Earth-2. It was also emphasized that not just the weights, but the training data and reproduction scripts are being made publicly available.
What I found interesting was that CEO Huang acknowledged that "NVIDIA's models are not exactly at the cutting edge of the frontier, especially in language." He then continued by saying that through post-training on your own data, you can turn the data you have on hand into sovereign intelligence.
The open models blog from the same day mentions that following Nemotron 3.5 Super VL and Bolt 74B on October 15, the smaller Bolt E4B 4GB is scheduled to be released on October 23. I ran both Super VL and Bolt 74B on my DGX Spark shortly after they were released.
The newly introduced knowledge element is NVIDIA Agent Knowledge Foundation. It is a collection of models and libraries for connecting internal documents and databases to agents, and includes Nemotron Retriever for document search and the new Kumo family for making predictions from tabular and time-series data.
These are bundled into the NVIDIA Agent Toolkit, and the keynote featured videos showing use cases in drug discovery, factories, and vehicle design.
Standardizing Physical AI with Hyperion and Agile ONE SN
The second half of the keynote focused on physical AI. CEO Huang reiterated the concept that three computers are needed: one for training, one for digital twins, and one to deploy on robots.
Examples shown included efforts to make communication networks an AI foundation, adaptive radiation therapy, and autonomous driving. For autonomous driving, Hyperion was promoted as the standard architecture, and adoption by Volvo Cars and others was announced.
For humanoid robots, the reference humanoid robot Agile ONE SN was announced. It is a 31-degree-of-freedom robot body made by Agile Robots, aimed at allowing researchers to avoid building both the hardware and software from scratch. Delivery is planned for early 2027.
Expanding the Local Development Environment with RTX Spark and DGX Spark
CEO Huang said that PCs going forward will not be mere tools but companion agents, and will need to have all 5 elements. He held up the RTX Spark, which went on sale October 16, and compared its 1 PFLOPS performance to the 2016 DGX-1, showing how it has become a size you can hold in your hand.
The same-day local AI blog summarizes the three local models.
| Model | Key details |
|---|---|
| RTX Spark | On sale October 16. Up to 128GB of unified memory |
| DGX Spark | 64GB version on sale October 23. Supports clusters of up to 8 units |
| DGX Station | 748GB of memory. Windows version expected in December |
As a DGX Spark user, software updates are of interest. NVIDIA Sync's Model Launcher will allow you to download models and containers and connect them to OpenCode. A collection of locally quantized checkpoints for Hugging Face has also been published.
Microduck, Introduced at the End, Was Said to Have "Been Born on RTX Spark"
The closing segment of the keynote was a demo of an agent on RTX Spark orchestrating robot training. The star was Microduck, a small bipedal robot made by Pollen Robotics, a Hugging Face subsidiary. The robot stands about 25 cm tall and is priced at $399.
Pollen Robotics also released the tabletop robot Reachy Mini last year, and I've purchased both. I ordered Microduck at the time of the announcement.
In the demo video, Hermes Agent and Nemotron plan a course on RTX Spark and build a simulation environment. Training runs in parallel across thousands of instances in the cloud, and the resulting policy is evaluated on RTX Spark before being distributed to the physical unit.
After the video, 4 Microducks appeared on stage. CEO Huang spoke to them saying "You were born on RTX Spark," spelled out the name Pollen for the parents, and called on the streaming audience to place orders too.
The keynote concluded with a closing video that recapped the content in song, with the Microducks remaining on stage.
Pollen's training repository and Isaac Lab also include walking tasks for Microduck, so you can try it in simulation even without the physical unit.
Summary
The GTC Berlin 2026 keynote reorganized the explanation of agents from Taipei in June into 5 elements: model, harness, sandbox, knowledge, and tools. It framed everything from AI factories to local PCs within that framework, and concluded by connecting agents and physical AI through the Microduck demo.
New announcements included the DSX POD, OpenShell Reference Architecture, Agent Knowledge Foundation and Kumo, and Agile ONE SN. Even for those building agents locally with DGX Spark or RTX Spark, it was a keynote with many immediately relevant topics, such as Model Launcher and the OpenShell reference configuration.
The three themes of OpenShell, open models, and sovereign AI—which underpinned the narrative directed at Europe—are all closely related to my own past research. I'd like to write separate articles on these topics, comparing them with my own hands-on experience.
Reference Links
- GTC Berlin 2026: Live Updates From NVIDIA CEO Jensen Huang's Keynote (2026-10-21, NVIDIA Blog)
- NVIDIA GTC Berlin: Live Updates on What's Next in AI — List of announcements
- NVIDIA Introduces DSX POD AI Factories to Accelerate Europe's Infrastructure Deployment (2026-10-21)
- European Ecosystem Builds World's Most Efficient AI Factories With NVIDIA DSX (2026-10-21)
- NVIDIA Launches OpenShell Reference Architecture for Safe AI Agents (2026-10-21)
- NVIDIA Open Models Advance Europe's Sovereign Intelligence Across Industries (2026-10-21)
- Open From Data Center to Desktop: NVIDIA Expands Nemotron Family With Highest Efficiency Super VL and Bolt Models (2026-10-15)
- NVIDIA Agent Knowledge Foundation
- NVIDIA Agent Toolkit Expands, Grounding Enterprise Agents in Company Data (2026-10-21)
- European Mobility Leaders Build Safe Vehicles and Robotaxi Fleets on NVIDIA Hyperion (2026-10-21)
- NVIDIA Advances Isaac Platform for General-Purpose Robotics With New Reference Humanoid, Frameworks, Simulation Pipelines and Data (2026-10-21)
- Bringing the Spark to Europe: Local Intelligence Powers Next Wave of Agents (2026-10-21)
- NVIDIA RTX Spark PCs Available Today — A New Beginning for Windows PCs (2026-10-16)
- pollen-robotics/microduck_rl
- isaac-sim/IsaacLab PR #8270 — Microduck walking task
- NVIDIA GTC Berlin 2026 Keynote With CEO Jensen Huang — Full keynote
- Official Keynote Opening Video | NVIDIA GTC Berlin 2026
- Official Keynote Closing Video | NVIDIA GTC Berlin 2026