GTC Berlin Keynote Recap: DSX POD, OpenShell & the Humanoid Robot

By Muzamil Ahmad · Updated 2026-10-21 · Productivity

#productivity

Save

Two days ago we published our GTC Berlin preview, laying out what Jensen Huang might actually announce at his October 21 keynote. The keynote has now happened — 11am to 1pm CEST at Berlin's Tempodrom, in front of a packed house — and the preview is officially history. This is what really happened, what the preview got right, and what nobody saw coming.

The short version: Huang unveiled DSX POD, factory-built AI data centers that ship in containers and deploy in 14 days; OpenShell, a reference architecture for keeping AI agents "contained, monitored and under control"; a 1.4-metre German-built humanoid robot he says is racing toward a "ChatGPT moment"; and a robotaxi deal with Bolt that puts 25,000 self-driving cars on NVIDIA's Hyperion platform. Oh, and there were robot ducks. And protesters.

Scorecard: What Our Preview Predicted vs. What Happened

The preview's honest caveat — "treat everything marked 'expected' as informed reading of the signals, not confirmed launches" — held up. But the keynote had more hardware-logistics news than anyone guessed.

DSX POD: The AI Factory That Ships in a Container

The biggest infrastructure announcement was DSX POD, a reference design for complete Vera Rubin AI factories that partners build, cable and test inside modular containers before delivery. Once the site utilities are ready, NVIDIA says on-site deployment shrinks from six months to as little as 14 days.

"This is a complete transformation of the world supply chain," Huang said on stage. "This will change the way that people install and build data centers, and hopefully, as a result, velocity goes up."

Each pod combines seven liquid-cooled Vera Rubin compute racks using NVIDIA's Spectrum-X networking. Seventeen partners — including Dell, HPE, Lenovo, Foxconn, Schneider Electric and defense company Anduril — plan to deliver DSX POD factories worldwide in early 2027.

The efficiency claims are aggressive: up to 100% less cooling water (closed-loop liquid cooling with coolant up to 45°C, plus dry coolers that release heat without evaporating water) and up to 30% less power for the same AI output via DSX MaxLPS software. Huang claimed that at 2 gigawatts, DSX designs can save 54 million liters of water a day. CoreWeave, Google Cloud, Lambda, Nebius and Nscale are among the companies planning to use the platform.

Honest caveat: the 14-day figure assumes a prepared site — power, connectivity and permits are somebody else's problem, and that is historically the part that takes the longest.

OpenShell: A Leash for AI Agents

With enterprises about to deploy AI agents by the billions, Huang's answer to "who watches the agents?" is hardware. The OpenShell Reference Architecture is a recipe for building, containing and monitoring AI agents:

• Models think on Vera Rubin systems
• Agent harnesses and tool use run on Vera CPUs
• BlueField-4 chips run NVIDIA's Sentry software, watching agent behavior from outside the main system and enforcing policy in hardware

"As enterprises prepare to deploy billions of AI agents, safety and control must be built into the computing architecture," Huang said. "Agents can think on Rubin, act on Vera and be governed on BlueField."

NVIDIA lists Anthropic, AWS, Google Cloud, Meta, Microsoft, OpenAI, Perplexity, SAP and SpaceX's AI unit among those working with the architecture, alongside security firms like CrowdStrike and Palo Alto Networks. Alongside it, NVIDIA expanded its Agent Toolkit with tools that connect agents to company documents and databases — already adopted by SAP, ServiceNow, Snowflake, Databricks and Palantir. "The most important thing is to make sure the agent is secure," Huang said.

Honest caveat: as The Register noted, the architecture requires NVIDIA hardware end to end, OpenShell still has open security issues on GitHub, and it collects telemetry by default unless users opt out. Banks BBVA and Santander are reportedly reviewing the spec — interested, not committed.

For context on what this trickle-down means for smaller companies, see our guide to AI tools for small businesses.

The Humanoid: "The ChatGPT Moment for Robotics Is Fast Approaching"

"The ChatGPT moment for robotics is fast approaching," Huang declared — and then showed the machine he is betting on. The Agile ONE SN, built by Munich startup Agile Robots (one of four German startups NVIDIA announced investments in two days earlier), is a reference humanoid meant as a starting point for developers and university labs rather than a finished product: a complete body, sensors and software stack that researchers can use to train and test new skills without designing their own hardware first. Huang said it should cut researchers' setup time from weeks to days.

The specs: 1.40 metres tall, 40 kg, 31 degrees of freedom, carrying 3 kg in each arm, with two five-fingered hands that have 20 degrees of freedom each plus touch sensing. It runs NVIDIA's Jetson Thor computer and Isaac GR00T robot models. Stanford University and ETH Zurich are reportedly interested in using it. It is due in early 2027; no price was given.

Agile Robots founder Zhaopeng Chen framed the bet this way: "The breakthrough for humanoids is usefulness at scale: one adaptable machine that learns many skills and works across environments built for people."

The robot arrived with a full software kit: Isaac Lab 3.0 robot-training software went generally available, a library of more than 2,000 simulation-ready robot and factory parts landed on Hugging Face, and NVIDIA released a dataset of human activity filmed from three angles for teaching robots by example. "With Nvidia Isaac, our full-stack open robotics platform, we are giving developers the models, data, simulation tools and computing systems to bring physical AI to the world's largest industries," Huang said.

One pointed detail: a German-made design now sits at the center of NVIDIA's robot platform in a market where Chinese companies built more than 95% of the world's humanoids in the first half of this year. That is not an accident — it is industrial policy with a Jetson Thor inside.

The on-stage warm-up act for all of this was a flock of Microducks from Pollen Robotics — small two-legged robots with cameras for faces that followed Huang around, lay down and rolled over on voice command. They demonstrated a workflow where AI was trained in the cloud, evaluated on an NVIDIA RTX Spark PC, then deployed onto the robots.

Robotaxis and Phone Masts

Two announcements pushed NVIDIA deeper into industries that have nothing to do with data centers.

Self-driving: ride-hailing company Bolt chose NVIDIA's Alpamayo self-driving software for its European robotaxis. The plan: start gathering driving data in Madrid in 2027 (subject to regulatory approval), then Germany, as part of a planned rollout of at least 25,000 self-driving Lucid cars built on NVIDIA's Hyperion platform. Huang boasted that Hyperion has become an industry standard, with next-generation vehicles from Mercedes-Benz, Volvo, JLR and Volkswagen's van division all compatible with it. Robotaxis remain rare in Europe — even Waymo's London launch keeps slipping — so this is a long bet, not an imminent service.

Telecoms: T-Mobile made what NVIDIA calls the world's first "AI-RAN" call on a commercial network — in Bellevue, Washington — using NVIDIA and Nokia equipment that runs AI inside the mobile network itself. BT, Deutsche Telekom and Elisa are running pilots, with BT planning live trials in Birmingham in 2027. "After working on this for nearly eight years, Nvidia made the first commercial phone call on the T-Mobile network with Nokia, and so now CUDA makes phone calls," Huang said. European telecom operators are also deploying NVIDIA-powered computing in their networks to deliver AI services — the network itself is becoming an AI customer.

The Money: Nearly $10 Billion for Europe

The keynote closed out the financial storyline the preview was tracking. Huang said the weeklong roadshow through Paris, London, Stockholm and Dublin brought NVIDIA's commitments to Europe's technology ecosystem to nearly $10 billion (about €8.6 billion) — roughly $8 billion deployed over the past year plus the ~$2 billion in new deals he promised in Paris on October 13. The company says it works with more than 12,000 partners across the continent, and made 14 startup investments during the trip alone. A $250 million computing pledge for UK AI safety research was also announced on the tour.

He opened the keynote by asking what L'Oréal, Revolut, Spotify, Merck, Vodafone, Zara, Porsche, Nestlé and HSBC have in common: they are all European companies working with NVIDIA on AI.

The Germany deal, signed

The industrial AI pact our preview tracked landed two days before the keynote. On October 19, Germany's Ministry of Economics signed a joint declaration of intent with NVIDIA to promote AI adoption in German industry, with about a dozen companies participating — including Siemens, Audi and Black Forest Labs (the lab behind FLUX).

That same day, Huang met Chancellor Friedrich Merz in Berlin. Merz made a direct appeal to the Mittelstand — Germany's family-owned manufacturers: "Please open your doors for AI to make use of your data." He said Germany wanted to be "one of the leading, if not the leading, country in the world" for AI adoption across manufacturers, machinery makers, carmakers and their suppliers. Huang pitched AI as augmentation, not replacement: "That knowledge belongs to you. It cannot be replaced. AI can only augment your capability, it will not replace your capability."

The reported terms: Germany is expected to put up several hundred million euros in government funding, much of it to rent computing power on Deutsche Telekom's AI cloud, built on NVIDIA's Nemotron language models and Cosmos world models. Honest caveat: the funding figure comes from Handelsblatt via government sources and has not been confirmed by either the government or NVIDIA. Full implementation details are still to come.

Also on October 19: NVIDIA announced planned investments in four German startups — Agile Robots (robot arms and the Agile ONE humanoid), Avelios Medical (an AI-native hospital information system), Mirelo (AI-generated sound for video) and Reflex Aerospace (satellites built in about 12 months). And NVIDIA researchers used Jülich's JUPITER supercomputer to run a record 104-qubit simulation of a fault-tolerant quantum computer.

The Framing: "A New Industrial Revolution"

Huang's rhetorical through-line: "We are now at the beginnings of a new industrial revolution." He described AI as a five-layer cake — energy, advanced computers, AI factories, tokens and digital intelligence, and applications on top — arguing the infrastructure must exist before the models can emerge, the way networking preceded the internet and the grid preceded electricity. "Long before the AI models emerge, the infrastructure has to exist."

The Europe pitch, repeated all week: "To be part of this AI economy, you don't want to just be consuming the AI; you want to be producing the AI." As Forrester analyst Paul Miller put it, Huang positioned NVIDIA as a strong supporter of Europe's sovereign tech ambitions — while noting the irony that countries buying independence from foreign tech providers are mostly buying it from one American chipmaker.

And the jobs question, answered with characteristic Huang confidence: "Some people think AI will mean the end of people who write code, but the opposite is true. More people will write programs using the new tools, and every company will become an AI company." He also claimed the processing capabilities of NVIDIA's AI chips have increased 1,000-fold since their introduction.

The Skeptical View

Not everything went smoothly. Environmental protesters briefly interrupted the keynote, holding up banners and demanding: "Why don't you clean up your act?" Huang handled it with humor — "All important issues, but let me get through my keynote. I'll talk to you guys later" — as security removed them. The water-usage claims for DSX (54 million liters a day saved at 2 gigawatts) landed moments later, which may or may not have been coincidence.

The bigger skepticism is financial. Axios noted the keynote lands at a "precarious moment" for the industry, with investors questioning whether massive AI spending converts into returns. Huang's answer was volume: more factories, more agents, more robots, more cars, more phone networks — demand for compute everywhere, forever. Whether the accounting catches up with the ambition is the story of the next year, not this keynote.

What Wasn't Announced

For the record: no new flagship GPU architecture. Vera Rubin was declared to be in full production, but the next chip family stayed in the lab. Our preview called this one exactly right — and DSX POD shows where the "hardware" announcements actually went: logistics and packaging, not silicon.

Also absent: any specific open-model launch from NVIDIA itself, despite the open-models track and Black Forest Labs presence. The SOOFI open-model consortium work surfaced only in the Germany-deal coverage.

Bottom Line

GTC Berlin 2026 was not a chip keynote. It was a deployment keynote: how fast you can stand up an AI factory (14 days), how safely you can run a billion agents (OpenShell), where the robots come from (Munich), and who pays for Europe's AI independence (nearly $10 billion, mostly NVIDIA's money). The preview got the money, the Germany deal, the agents, the robots and the no-new-GPU call right — and missed the containers, the robotaxis and the phone calls.

The honest verdict: it mattered. Not because of any single product, but because it showed NVIDIA's strategy in full — sell the chips, then sell the factory, then sell the safety system, then sell the robot that runs on all three. Every layer of the five-layer cake, with NVIDIA taking a cut at each one.

Our original preview covers the run-up to the keynote — the Europe tour, the Germany deal reporting, and what was expected. Keynote coverage draws on The Register's reporting and a detailed recap of the announcements.