Tempodrom Address Maps Sovereign Compute Buildout For Regional Industry

The evening opens inside Tempodrom, the round arena in Berlin that hosted the 20-22 October GTC gathering, with its program of more than 100 sessions. The keynote itself took place on 21 October before a European crowd. Early on, the speaker turned the spotlight outward, asking which regional companies the audience wanted to see lead: L'Oreal, Revolut, Spotify, Merck, Vodafone, Zara, Porsche, Nestle and HSBC all got a mention. The roll call framed the night as a conversation with Europe rather than a broadcast to it.

That framing continued with a recap of a tour through Paris, London, Stockholm and Dublin, presented as the groundwork for a long-term regional commitment of about $10B. Nvidia live blog notes describe the same tour as the launchpad for expanded developer outreach across the continent. The figures quoted were 12,000-plus partners and 14 startup investments. Together they set up the central pitch: Europe should build its own intelligent infrastructure instead of renting it from elsewhere.

The conceptual core arrived as a five-layer cake: energy at the base, then computing capacity, then the AI factory, then tokens, then models, then applications at the top. Each layer converts the one below into something more valuable, with electricity ultimately refined into intelligence. The memorable line was rephrased rather than repeated: stop merely consuming outputs and start producing them. It is an industrial metaphor, casting inference as manufacturing and the audience as prospective factory owners.

From metaphor the talk moved to the full technology stack behind an AI factory: hardware systems, system software, simulation engines grouped under CUDA-X, and foundation models. The argument was that factories need every layer co-designed, not assembled from spare parts. Simulation and solver engines were presented less as research tools and more as working equipment for agents. The throughline was vertical integration, from power intake to application output.

The hardware headline was Vera Rubin entering full production, embodied by the NVL72 rack in which 72 graphics processors operate as a single giant processor. Nvidia newsroom reporting places agentic throughput at 10x the level of the Grace Blackwell generation. Volume manufacturing, rather than a lab prototype, was the point being stressed. Rubin supplies the reasoning muscle while companion systems handle interaction, a division of labor revisited later in the evening.

Interconnect got its own beat: MVLink switching ties the 72 processors together inside that single-processor illusion. Assembly that once took two hours now takes five minutes, a compression meant to show factory-style manufacturing discipline. Cooling is handled entirely with liquid, removing air-cooled bottlenecks from dense racks. The message was density without drama: more output per rack, less floor space per unit of intelligence.

The infrastructure announcement was DSX POD, a modular container-based factory that cuts on-site setup from six months to 14 days. Nvidia newsroom coverage puts that 14-day figure at the center of the European deployment story. Power efficiency improves too, with the same output available on up to 30 percent less draw, while closed-loop dry cooling can cut water use by up to 100 percent. For regions where permits and grid connections stall projects, the pitch is speed: arrive with containers, leave with a running factory.

Ecosystem names followed: Schneider Electric, Vertiv, Dell, HPE, Lenovo and Bull were presented as rollout partners, with first deliveries expected in early 2027. NextPlatform reports that a 1 GW datacenter costs about $50B to build while promising roughly $40B in yearly revenue, so a three-month delay can erase nearly $10B. That arithmetic is the commercial engine behind modular deployment: every week saved is measurable money. The segment closed the loop between engineering metrics and balance-sheet urgency.

Midway through, the focus shifted from factories to their inhabitants: agents were declared the new computing platform. Five ingredients define the recipe: a foundation model, a harness for orchestration plus memory, a sandbox for isolation, knowledge sources, and tools for acting on the world. Each ingredient maps to a failure mode it prevents, from forgetfulness to unsafe actions. The framing deliberately echoes earlier platform shifts, positioning agents as the layer where developers will now build.

OpenShell is the proposed runtime for that platform: an open-source, policy-governed layer with kernel-level sandboxing for agent workloads. The OpenShell repo on GitHub lays out four building blocks called Gateway, Supervisor, Sandbox and Policy prover. Policies travel with the workload, so permissions are checked continuously rather than once at launch. Thinking may run on Rubin-class systems while actions execute on Vera-class CPUs under BlueField supervision, keeping power and oversight in separate hands.

A reference design ties the pieces together: Rubin for reasoning, Vera CPUs with up to 2x prior responsiveness for interaction, and BlueField-4 plus Sentry for millisecond quarantine of suspicious behavior. Developer notes from Nvidia describe how the same governed pattern scales across production agent fleets. CoreWeave numbers illustrated the payoff: 3x faster sandbox startup, 1.8x agent loop speed, and more than 11,000 concurrent environments in a single rack. Named adopters spanned Anthropic, AWS, Google Cloud, Meta, Microsoft, OpenAI, SAP, Siemens, the SpaceX Super Intelligence team, CrowdStrike and Perplexity.

Open weights received equal billing as the route to sovereign capability: Nemotron for language, Cosmos for world understanding, GR00T for humanoids, Alpamayo for autonomy, Ising for quantum-classical crossover, and Earth-2 for physics. Nvidia product pages frame post-training as the step that turns these shared foundations into national capability. The doctrine is that intelligence becomes sovereign only when a country or company can adapt the substrate itself. Openness here means leverage, not charity.

Agents also need current knowledge, which arrived as its own layer: an Agent Knowledge Foundation plus Kumo for structured forecasting over enterprise data. Nemotron-based parsing, embedding and reranking prepare documents for retrieval at answer time. The claim is that fresh context matters more than frozen parameters for operational questions. Knowledge, in this telling, is infrastructure rather than an afterthought.

Developers got a number to take home: more than 11M builders work with CUDA-X libraries today. Simulation, solver and data engines were recast as agent tooling, giving software teams physics-aware instruments. The ecosystem argument is that hardware without programmers stays idle, so toolchains are the real moat. Eleven million is both a milestone and a recruiting poster.

Telecommunications offered the most concrete deployment story: roughly 100 thousand field sites represent about 100 GW of distributed capacity waiting to host inference. A first commercial call placed jointly across T-Mobile and Nokia equipment demonstrated voice running through accelerated infrastructure. The quip of the evening played on the verb: CUDA now places phone calls. Carriers, in this vision, convert towers into revenue-generating compute.

Healthcare and mobility shared the applied section. Siemens Healthineers showed adaptive radiotherapy in which Nemotron plans treatments, Omniverse twins the suite as a digital double, and Holoscan with IGX Thor adjusts delivery live. On roads, the Hyperion platform named Bolt with Lucid, Mercedes-Benz, VW Commercial, Einride, Volvo on the Alpamayo Stack, and JLR. Two industries, one pattern: validated autonomy under expert supervision.

Physical AI closed the technical arc with three computers for every robot: one to train, one to simulate, one to run onboard. Agile ONE SN, a reference humanoid with 31 degrees of freedom due next year, embodied the training target. The Isaac platform structures development through Capture, Lab and OS stages, while Cosmos 3 arrives in Super 64B, Nano 16B and Edge 4B sizes. Omniverse hosts the simulation leg, letting fleets rehearse before touching real floors.

The finale lightened the mood: RTX Spark, a palm-sized unit with 1 petaflop and 128GB matching a 2016 DGX-1, plus four Microduck robots from Hugging Face trotting on stage for a singalong-styled sendoff. Greenpeace demonstrators briefly interrupted proceedings over environmental claims and drone allegations, answered with a call to stay calm and talk later. Taken as a whole, the evening was a construction schedule disguised as a keynote: concrete dates, named partners and testable numbers rather than distant promises. The open question is execution, and the milestones above give observers plenty to verify.