Huang Tells Europe to Accelerate AI on Nvidia’s Terms

Nvidia chief Jensen Huang used a Berlin keynote on October 21, 2026, to tell Europe to adopt AI faster across every industry. He had just come from Paris, London and Stockholm with the same brief.

The product under the speech is a containerized Nvidia factory. The independence Europe is funding still has to plug into his stack, and into grids that ration new load.

Huang Spent a Week Telling Europe to Stop Overthinking

Huang opened GTC Berlin at the Tempodrom on October 21, 2026, after stops in France, Sweden, the United Kingdom and Germany, and told the room that every industry had to become an AI company if Europe wanted an AI economy of its own. He named cars, banks and hospitals as the buyers who still had to move.

“For Europe to benefit from AI, and to have really created an AI ecosystem and AI economy, you need all the end markets, the end industries to adopt AI,” he said. The next wave, he added, is physical: robots, plants, vehicles and heavy industry. “Electrons will never replace atoms.”

The tone was already set before the keynote. In London on October 14 he told reporters the country had the universities and the talent and still talked itself out of using them. “If there was a million ways to win, you find the one reason not to do it,” he said. “I think you overthink things.” He told founders to put the pedal to the metal. In Berlin he carried the same line into a room with Chancellor Friedrich Merz.

THE WEEK HE WORKED THE CAPITALS

1. October 13, 2026: Meets French officials in Paris and names new bets on Doctolib, Entalpic, Genesis AI and Pigment.
2. October 14, 2026: Tells London the United Kingdom overthinks failure, and pledges $250 million of compute time over five years for agent safety work, including at Oxford and University College London.
3. October 16, 2026: In Stockholm, at the launch of the Sferical One supercomputer, says every company will be an AI company and that Europe has to build the full stack, from halls of servers to the apps on top.
4. October 19, 2026: Sits with Merz at Kraftwerk, a former power station, as Germany and Nvidia back physical AI for factories and cars.
5. October 21, 2026: Delivers the GTC keynote at the Tempodrom, 11 a.m. to 1 p.m. CEST, and unveils a boxed factory Nvidia says can go live in 14 days on a site that is already wired.

He also kept repeating a narrower claim about safety: regulate the applications, not the underlying technology, and do it now. That is a vendor’s map of the law. It is also the map that puts Nvidia’s hardware on the unregulated side of the line.

What Sovereign AI Means When Nvidia Sells the Chips

Huang has spent years selling “sovereign AI,” the idea that a country should train and run models on its own soil instead of renting them from the United States or China. In Berlin he put a sharper point on it. National know-how in plants and cars cannot be replaced, he told Merz, and should not be shipped out to be sold back as someone else’s model.

“That knowledge belongs to you,” Huang said. “It cannot be replaced. AI can only augment your capability, it will not replace your capability. Take advantage of the technology. It’s a new instrument.”

The instrument, in this telling, is imported. Europe keeps the data and the shop-floor skill. Nvidia keeps the accelerators, the networking, the software stack and the reference design for the building that holds them. That is independence of a sort. It is also a deeper hold on the layer that turns electricity into tokens.

Brussels is trying to loosen that hold from another direction. The European Chips Act took effect on September 21, 2023, and the Commission in June 2026 proposed Chips Act 2.0 to cut chip dependence. Eighteen first-of-a-kind plants have already won state-aid clearance, for public and private outlays of more than €32 billion, including the ESMC venture in Dresden with TSMC, Bosch, Infineon and NXP. Those files are about making more silicon in the Union. Huang’s tour is about filling European halls with silicon designed in Santa Clara.

The AI Act is a thinner brake than the speech implied. Rules for general-purpose model providers have been in force since August 2, 2025, and the Commission has been able to enforce them, including with fines, since August 2, 2026. The Digital Omnibus that took effect on July 27, 2026, moved the heavy high-risk duties for stand-alone systems to December 2, 2027. Europe did already ease the timeline. Huang still talked as if caution were the only setting.

There’s nothing to regulate if the industry doesn’t thrive.

Jensen Huang, Nvidia CEO, to Chancellor Friedrich Merz in Berlin

Merz answered that Europe was still hunting for a balance, and that neither a free-for-all nor a clamp would do. Huang’s reply was a sales close dressed as patriotism: give Germany a chance, give this industry a chance.

Germany Signed a Physical AI Pact in a Power Plant

On October 19, in the hall of Kraftwerk, Merz, the Federal Ministry for Economic Affairs and Energy and the Federal Ministry for Research, Technology and Space signed a joint declaration of intent with Nvidia. The text backs a joint push on models for physical AI, using German industrial data, and ties Nvidia more tightly to the government’s Industry 4.0 platform.

Merz told the room Germany wanted to be among the most advanced places in the world at putting AI into manufacturing, machinery, cars and the supplier base. He turned to the Mittelstand and asked those firms to open their doors and use the data they already hold. That is the part of the pact that is not a photograph. Factory data either moves onto Nvidia’s tools or it does not.

Huang praised German engineering and called the shift a new industrial revolution. “In the Chancellor’s words, you are re-engineering Germany. This is your chance to do it well.” Merz said he wanted Huang in his office at least once a year, “extremely happy about Germany as a place to invest.”

The official account posted the meeting the same evening, with Huang and Merz side by side and four German firms named as new planned investments.

Germany knows how to build, and now it’s bringing that expertise to industrial AI. 🇩🇪

Tonight in Berlin, @JensenHuang and Chancellor @_FriedrichMerz discussed combining German engineering with NVIDIA AI to advance manufacturing and automotive.

We’re also backing Germany’s… pic.twitter.com/9uwH0rYYLk

— NVIDIA (@nvidia) October 19, 2026

Siemens, already in a deeper factory tie-up with Nvidia, used the same day to talk about its Erlangen plant as a test bed. Nvidia says virtual production ramp-up there can cut engineering effort by up to 85 percent, and that agent-based planning can cut costs by up to 20 percent. Those are Nvidia’s figures, from a partner that still has to buy the boxes.

Berlin treated Huang as a co-author of industrial policy, not a chip vendor dropping in for a photo. The tax-and-welfare objection that followed online is the prosaic version of the same fact: the state is rolling out the carpet for a company whose pricing power is the point of the visit.

Speed to Plug In Still Trails Speed to Buy Chips

A September 2026 brief from the Brussels think tank Bruegel put the Union at two gigawatts of AI compute, 5 percent of the world total. The United States holds about 35 gigawatts, or 78 percent. China holds about 5 gigawatts, or 11 percent. The world total in that tally is 45 gigawatts.

AI COMPUTE IN 2026

Bruegel’s authors say private capital is not the binding limit. Speed to operation and grid access are. Their brief notes that the Commission’s proposed Cloud and AI Development Act tries to hurry permits and does not cover grid connections. In Germany, they put the average wait at 42 months.

In March 2026, Denmark’s grid operator Energinet paused new large-scale connection deals after about 60 gigawatts of requests stacked up against peak demand of about 7 gigawatts. Ireland, the Netherlands and parts of Germany have already rationed or conditioned big new loads. Huang can fly four capitals in a week. A substation cannot.

The Union is trying to buy its way around the gap with public compute. Bruegel describes an AI Gigafactories plan that would put €20 billion of EU money against as much as €200 billion of private spend, for up to five giant training sites, plus smaller “AI factories” already being stood up on EuroHPC machines. Those sites still need power in the same grids, and they still need accelerators. Most of the planned capacity in Bruegel’s project list is privately financed. The chips inside the racks are still, in the main, Nvidia’s.

A Containerized Factory Promises First Tokens in 14 Days

The hardware answer Huang brought to Berlin is the DSX POD, a partner-built reference design for a full Vera Rubin AI factory packed into modular containers, tested at the plant, then shipped. Nvidia says that at a prepared site, with power, cooling and networking already in place, the design cuts time to first production from six months to 14 days. Deliveries are aimed at early 2027. Partners on the list include Bull, Dell, HPE, Lenovo, Schneider Electric and Vertiv.

THE DSX POD PITCH

• The clock: Six months of on-site build falls to as little as 14 days once the site is already wired for power, cooling and network.
• The caveat: The box does not create a grid connection, a water permit or a plot; it only fills a site that already has them.
• The water claim: Closed-loop liquid cooling and dry coolers can cut facility cooling water use by up to 100 percent, Nvidia says, with a modelled 2-gigawatt deployment saving 54 million liters a day, equal to about 180,000 households.
• The power claim: The same output with up to 30 percent less electricity, and tools to flex load with the grid.

“An AI factory creates value when it starts producing tokens, not when construction is complete,” Huang said. He is right about idle capital. He is also selling a stopwatch that starts after the hard European work, the substation and the permit, is already done.

That is why the 14-day figure and the 42-month German permit sit in the same story without cancelling each other. One is the time to rack a prepared pad. The other is the time to be allowed to draw current. Europe can adopt the box and still wait on the wire.

Nearly $10 Billion Buys Access Across Four Capitals

Nvidia put its European spend at nearly $10 billion across companies, universities and labs. The tour was the closing argument for that cheque: more local firms on the stack, more ministers on the stage, more reasons for a plant or a hospital to buy the next rack.

THE BETS NAMED ON THE TRIP

• France: New planned investments in Doctolib, Entalpic, Genesis AI and Pigment, after earlier cheques for Alice&Bob, Gradium and Scintil Photonics.
• Germany: New planned investments in Agile Robots, Avelios Medical, Mirelo and Reflex Aerospace, on top of n8n, Neura Robotics and Black Forest Labs.
• United Kingdom: Stakes in startups Huang toasted in Mayfair with AI minister Kanishka Narayan, plus $250 million of compute time over five years for safety research, and a public push behind the British cloud firm Nscale.
• Sweden: A national supercomputer, Sferical One, kept on Swedish soil with Nvidia chips, sold as a way to hold sensitive data at home.

None of those firms makes the accelerators. They make robots, hospital software, audio models, satellites, clouds and clinic apps. That is the European “economy” Huang is offering: local software on his silicon. Startups that want to remain based in Europe still need local capital, which is why a fund built to stay with European AI firms keeps writing cheques inside the region. Nvidia’s money pulls in the other direction, toward a stack whose scarce part is not the application.

In London he also said he would, in a hypothetical, cut off chips to a customer building unsafe systems, after a conversation. That is a useful line for a safety audience. It is not a substitute for a second supplier.

Europe’s Compute Share Barely Budges After the Factories

Bruegel’s 2031 forecast is the cold arithmetic under the applause. Even with the projects now listed, including the gigafactory plan, the Union reaches 20.89 gigawatts, or 5.6 percent of a world total of 373 gigawatts. The share barely moves. The United States is still expected to hold most of the world’s AI compute. China gains some. Europe remains a rounding error with excellent universities.

Huang’s week does not change that math. It changes who Europe has to call to try. Merz wants a yearly meeting with a happy investor. The Mittelstand is being asked to hand factory data to tools that sit on Nvidia’s platform. The DSX POD will shave months off a fit-out for anyone who already has a live grid slot. Energinet, TenneT and the German distributors are still the people who decide whether the slot exists.

Europe can go faster on applications, as Huang asks, and still deepen the dependence its Chips Act 2.0 is supposed to shrink. It can also keep the data at home, as he advises, and still import the engine. The guardrails he wants are on software. The guardrails that actually slow him are made of copper, permits and peak load.

Merz already booked the follow-up. He wants Huang back in the chancellery, at least once a year, to measure whether Germany became a place that made the world’s most valuable chip firm glad it came.