Europe’s AI Strategy Depends on Building Local Compute and AI Factories

Jensen Huang framed the keynote around a shift in what computers do. AI, in his account, is not merely software running on a computer people operate. It is a new kind of industrial infrastructure: factories that generate tokens—“the building blocks of AI”—and supply intelligence to applications and industries.

He described this as a new industrial revolution because AI changes both how things are built and what can be built. It can become part of every company’s operations, he argued, while enabling products and industries that did not previously exist. His analogy was electricity: AI is digital intelligence that can be used across sectors, but making it useful at scale requires infrastructure to reach the places where work happens.

That distinction leads to his central prescription for Europe: don’t stop at adopting AI services. Build the capacity to produce AI. Huang presented the opportunity as a five-layer stack: energy at the base, then compute, AI factories, models, and applications. The layers are connected, but each has its own companies, investment needs, and potential for growth.

The infrastructure must come first. Models cannot do useful work at scale without computers and networks to run them, just as the internet required networks and electricity required power plants and electrified buildings. Above that infrastructure, models extend beyond language to biology, chemistry, physics, robotics, and animation. Different industries, Huang said, will use different model types, which will in turn support new applications and AI-native companies.

He pointed to European firms across the stack, including energy and industrial companies, cloud providers, AI startups, and application developers. NVIDIA’s slide put the scale of its European activity at nearly $10 billion invested in the previous year, 14 startup investments announced during the week of the keynote, and more than 12,000 companies in its European ecosystem. Huang said the company had worked in Europe for 33 years, tracing its history there to early chips manufactured in France and assembled in Malta, a German workstation partner, and later work with Audi and Mercedes-Benz.

The historical thread served a strategic point: Europe is not starting from zero. Huang argued that the region already has industrial capabilities and relationships across the AI economy, and that the next step is to build more of the infrastructure and production capacity locally. In his framing, a European AI economy means more than access to models. It means energy, computing, factories, models, and applications developing together.

Huang’s “five-layer cake” is also a claim about what an AI company has to build. NVIDIA, he said, has moved from being understood primarily as a GPU supplier to offering a full-stack AI-factory platform: computing, networking, storage, system software, libraries, and models.

A large AI factory is a different kind of computer from the personal or enterprise computers people are accustomed to using. Huang described a one-gigawatt data center as an approximately $50 billion project, with land, power, buildings, compute, networking, storage, cooling, and software all needing to work together. The factory’s purpose is not simply to run programs on demand; it is to produce tokens continuously.

That scale changes the role of software. The factory requires system software to coordinate and operate its components, along with libraries that let developers use the underlying hardware for different kinds of computation. Huang described NVIDIA as an algorithm company that designs silicon to accelerate algorithms. The company’s CUDA-X libraries, he said, span areas such as computational lithography, signal and image processing, data processing, particles, fluids, and sparse solvers. He called those libraries the company’s “crown jewels,” arguing that they connect its computing architecture to the algorithms and industries that can use it.

He singled out cuDNN, NVIDIA’s library for deep learning, as foundational to modern AI. Huang compared its role in AI to SQL’s role in databases: a domain-specific library that made a field of computing accessible and productive. He also placed today’s models in the software stack as a new engine for applications, working alongside those libraries rather than replacing them.

The distinction matters to his argument about AI and existing software businesses. AI, he said, will not eliminate tools companies. It will give those tools more users and make the companies that provide them into AI companies. A design, engineering, or research tool can be paired with an agent that knows how to use it, rather than discarded in favor of a general-purpose chatbot.

This is one reason Huang emphasizes breadth. NVIDIA’s architecture is intended to support many workloads, not only one model type or algorithm. The same general-purpose platform can serve cloud providers, AI labs, neoclouds, AI-native startups, computer manufacturers, and industrial companies. Huang described the architecture as “productive, fungible, and durable”: productive because it should be used as soon as infrastructure is deployed; fungible because it can serve multiple workloads, industries, and stages of AI development; and durable because software updates and a large installed base can extend its usefulness.

For Europe, he linked that breadth to the need for local infrastructure and adoption. Cloud providers matter, but they are not the whole market. AI labs and neoclouds can build capacity within regions; original equipment manufacturers can build systems for enterprises; and end industries—from automotive and manufacturing to health care and financial services—must adopt AI if a broader AI economy is to emerge.

Huang’s account of the latest shift in AI demand begins with agents: systems that can reason, iterate, use tools, and perform tasks rather than only respond to a prompt. Because they take multiple steps and call tools, agents consume more tokens than simpler interactions. A chart attributed to OpenRouter showed agents using 34 times more tokens than at the start of February. Huang argued that more useful agents lead to more use, more token demand, and greater need for computing.

He broke an agent into five parts: a model, a harness, a sandbox, knowledge, and tools. The model handles perception, reasoning, and planning. The harness coordinates the work and gives the agent a structure for acting. The sandbox isolates the agent. Knowledge supplies information it needs, including company-specific data. Tools let it carry out work.

This framework changes what the underlying infrastructure has to process. A data center built only to run large language models does not necessarily address the full agent workload. The model must think, but the agent also needs a runtime to manage tasks and memory, computing for data and knowledge, and computing for the tools it uses.

NVIDIA’s Vera Rubin system is designed around that larger workload. Huang described a rack-scale system with separate components for model processing, secure agent processing, knowledge serving, and CUDA-X tools processing. The shift continues a progression from individual GPUs to whole systems, then to racks designed as a single large computer. With Vera Rubin, he said, the system is intended to process the agent as a whole rather than only the model at its center.

He presented the hardware as a series of linked design changes. The Vera Rubin system is fully liquid-cooled, with no cables or wires in the computer he demonstrated. Huang contrasted it with Grace Blackwell, which he said took about two hours to assemble and was 80 percent liquid-cooled; the new system, he said, takes less than five minutes to assemble. An NVLink switch connects 72 GPUs into one system. Each GPU, according to Huang, delivers 20 petaflops.

He also demonstrated Spectrum-X, describing it as an Ethernet switch using co-packaged optics. Rather than route data through a conventional circuit board and transceiver path, he said, optical components connect directly to the chip. The system he showed was described as a 512-port switch operating at 200 gigabits per second and in volume production.

Huang presented these components as parts of a single system rather than independent product announcements. The models process an agent’s reasoning; the agent runtime coordinates its actions; storage and data-processing components serve knowledge; and accelerated computing runs the tools. The point is not simply that each component is faster, but that the architecture is being organized around the complete agent workload.

He claimed Vera Rubin delivers another 30-fold performance increase over Blackwell, which he said was 30 times faster than Hopper for large-language-model processing. A visual attributed to SemiAnalysis showed a 30-fold reduction in the cost of generating tokens compared with the next-leading accelerator. Huang’s broader performance argument was that the gains come from “extreme co-design”: simultaneous work across chips, systems, software, algorithms, and the AI factory, rather than relying on improvements in processor design alone.

He compared this approach with Moore’s Law. Over a decade, he said, Moore’s Law would have implied about a 100-fold improvement, while NVIDIA’s combined design work had improved AI performance by a million times. The mechanism he emphasized was coordinated optimization across the stack.

The scale of an AI factory makes deployment time an economic variable. Huang said a one-gigawatt data center costs about $50 billion and could generate tens of billions of dollars in annual revenue once operating. If equipment sits idle while a site is assembled and commissioned, the delay postpones the output expected from a very large capital investment.

NVIDIA’s answer is the Vera Rubin DSX POD, a modular, factory-built AI-factory reference architecture. The idea is to assemble and test more of the system inside the supply chain, then ship a functioning unit that needs less on-site cabling and integration. The announcement slide described a factory-built, integrated, pre-tested cluster, with delivery-to-deployment cut from months to days. Huang separately said the interval from the arrival of computing to production could fall from four or five months to about a month.

He highlighted the amount of networking involved in these builds, citing 17,000 kilometers of wiring for Spectrum-X at the relevant scale. Moving more integration into the factory, he argued, could cut on-site assembly and reduce idle computing capacity.

This is both a manufacturing and a utilization argument. Huang said NVIDIA had already moved rack-scale systems into the supply chain, enabling production at a scale large enough to support hundreds of data-center build-outs around the world at once. The next step is to move more of the assembled AI factory into the same process. Faster deployment, in his account, means that expensive infrastructure begins producing tokens sooner and that less capital sits unused.

Huang illustrated the financial stakes with a hypothetical factory producing $40 billion in annual revenue. He initially said a three-month delay would mean $5 billion in revenue, then corrected himself to $10 billion. His point was that a short delay matters when the infrastructure is this expensive and the expected annual revenue is this large.

Moving integration into the supply chain is intended to reduce several kinds of delay at once: assembling racks on site, connecting extensive cabling, and qualifying the system after it arrives. Huang said the equipment could be run, checked, and burned in before shipment, so it would arrive functional rather than requiring months of assembly. He also argued that reducing the time between installation and operation would reduce idle computing and “wasted idle capex.”

The proposal shifts some complexity away from the data-center site and into manufacturing. The computer is no longer just a rack delivered for local assembly; it becomes a tested, integrated unit whose components and software are expected to work together before delivery. That places supply-chain coordination alongside chip performance as part of the system’s economics.

An agent is useful because it can act autonomously, but that same autonomy makes its permissions consequential. Huang said agents may have access to files, networks, resources, and tools, so the privileges they receive must be limited and monitored. Asking a model to behave safely is not enough.

NVIDIA OpenShell is intended to provide a secure runtime for agents. The accompanying demonstration described a system in which an operator defines what an agent can access, and OpenShell enforces those limits at the level of system calls, files, and network connections. A policy prover checks proposed policy changes against a permissions ceiling so that an agent does not acquire more access than intended.

The architecture separates the agent from the controls around it. Each agent runs in an isolated sandbox. Network requests pass through a supervisor outside the sandbox, where they can be checked against policy. Credentials remain with the supervisor and are added only when a request goes to an approved endpoint. The demonstration’s security claim was straightforward: an agent cannot expose a key it never possessed.

OpenShell combines Linux kernel isolation with runtime policy enforcement. NVIDIA Sentry adds an optional hardware-isolated monitoring layer running on BlueField-4. In the demonstration, Sentry monitors agent activity and can intervene if the agent appears to be planning an unintended action, cutting its connection to the model and quarantining it. The system also records tamper-evident evidence for operators to review.

Huang described OpenShell as open source under Apache 2.0 and said it runs from edge to cloud. He said more than 100 partners were advancing the platform, including frontier labs, cloud and operating-system companies, cybersecurity firms, and enterprise software providers. He also announced a partnership with Anthropic to create Claude OpenShell agents.

The security model is not presented as a single fixed permission set. Huang stressed that each task, user, and agent launch may require different privileges. An agent handling one person’s work may need different access from the same agent handling another task. Permission-setting, enforcement, and monitoring therefore have to be dynamic.

OpenShell addresses what the product demonstration called the “please don’t” approach to security—telling an agent not to delete files, push code to a main branch, or touch SSH keys. The demonstration argued that security must be engineered and enforced beyond the model’s own instructions. For companies deciding whether agents can be deployed inside their systems, the boundary around what an agent can touch is part of the agent architecture, not an optional afterthought.

Huang argued that companies will use a mix of closed, open, and custom models. General-purpose work may be handled well by broad models, but a company’s specialized work often depends on proprietary knowledge. In that case, a model can be adapted to the company’s data and processes rather than treated as a finished, one-size-fits-all product.

NVIDIA’s open-model portfolio spans several domains: Nemotron for language, Cosmos for physical-world understanding, GR00T for humanoid robotics, Alpamayo for autonomous navigation, Ising for quantum computing, and Earth-2 for physics AI. Huang said NVIDIA provides not only model weights but also associated data and reproduction scripts, so users can inspect the model’s provenance and recreate it.

He presented customization as a way to turn an open model into a specialized one. Slides showed examples involving AI21 and Alice, Baseten and CodeRabbit, Trajectory and Harvey, Scale and Skydance, and Palantir with Tampa General Hospital. The slides displayed point increases in accuracy over named comparison models: +41, +40, +30, +15, and +20 across the examples. Huang’s point was that adapting Nemotron with a partner’s data could improve performance on that partner’s task.

He described the work as a practical route to what he called “sovereign intelligence”: models shaped by an organization’s own data and expertise. The ability to adapt the model matters, but a model alone does not give an agent access to useful company knowledge. That requires a pipeline for turning data into information the agent can retrieve, interpret, and use.

NVIDIA’s Agent Knowledge Foundation is meant to provide that pipeline. For unstructured material, the diagram showed Nemotron Parse and Nemotron Embed, followed by cuVS for vector search and Nemotron Rerank for ranking retrieved material. For structured data, it showed Auto-Ontology dividing work into analytical processing with cuDF and predictive work with Kumo.

Huang’s example was a company trying to anticipate late or missing orders. A predictive model could identify likely delays; a search system could then locate an engineering document describing a replacement component. The knowledge foundation combines document understanding, database analysis, prediction, and retrieval, so an agent can move from recognizing a problem to finding relevant information for a response.

This is where the company’s older software investments meet its newer model work. Huang pointed to CUDA-X as a collection of algorithms and libraries used in scientific discovery, engineering, research, design, and creative work. The libraries are not decorative support for agents: they let agents call domain-specific tools that can process data, search repositories, simulate systems, and optimize decisions.

For Huang, the decisive difference between an agent that sounds capable and one that can do real work is access to tools. A scientist needs a laboratory; an engineer needs simulators and solvers. An agent needs the equivalent. A model can reason and plan, but it needs tools to test a hypothesis, operate on data, or act within a system.

The CUDA-X library family supplies many of those tools. Huang’s examples included Earth-2 for weather and climate modeling, Warp for physics simulation, cuOpt for route optimization, cuDF for data processing, cuML for machine learning, cuQuantum and NetKet for quantum computing, and Aerial for telecommunications. A video demonstration showed simulations of fluid dynamics, robotics, aerodynamics, factories, autonomous vehicles, weather, and Earth systems; it explicitly identified the footage as computer graphics and simulation, not generative AI.

Huang said these tools can run on the same accelerated-computing platform and perform certain workloads 10, 50, or 100 times faster than running them natively on a CPU. The larger claim is that an agent can use an existing body of specialized software rather than re-create its capabilities in a language model. As agents become responsible for more complicated tasks, their usefulness will depend on the available toolchain.

The keynote offered three examples of that structure in practice. Lila Sciences is building lab-assistant agents using BioNeMo and CUDA-X tools. Siemens is building factory agents using Cosmos and tools for planning and digital-twin work. Dassault Systèmes is building virtual companions using Nemotron and design and simulation tools. Each uses the agent components described earlier, but the specialized work differs.

In the Lila Sciences example, the goal is to design a therapy for B-cell lymphoma. The Lila Iris agent turns that goal into an experimental plan. In a computational “dry lab,” Lila agents use models for protein design and RNA optimization and run millions of virtual experiments. Promising candidates move into a science factory where robots and instruments make and test them; the results then return to the agent and inform further experiments. The work is represented as a loop between computational design, physical testing, and renewed planning.

That loop connects model output to laboratory work without treating simulation as the final answer. The computational stage generates and filters candidates; the physical stage makes and tests promising ones; and the results become input to the next round of design. The agent coordinates steps, while models, instruments, and researchers’ experimental setting provide the means to carry them out.

At Siemens’ Electronic Factory Erlangen, agents help bring products into production. Process-preparation agents use Cosmos to generate virtual prototypes and synthetic data, then train and validate inspection models before production begins. Intelligence Center X agents use orders, supplies, and line setups to recommend production plans. Digital Twin Composer checks those plans against factory data and physics simulation before changes reach the factory floor.

This example separates planning from validation. An agent can recommend how to arrange production using information about materials and line availability, but the recommendation is checked against a digital representation of the factory before a real-world change is made. The keynote’s emphasis was on using simulation and current factory data as part of the operating process, not simply generating a plan and applying it.

The Dassault example follows an engineer adapting a Volkswagen Golf door design for a Tiguan. Aura and Leo agents reason over engineering requirements; Aura plans the work around the vehicle’s dimensions, weight limits, and other constraints. A design sub-agent selects a compatible door and reshapes it in CATIA. A simulation sub-agent tests variants in SIMULIA, accelerated by CUDA-X libraries, and identifies a vibration problem. The result returns to the design agent for revision, after which an engineer reviews and selects a final design.

The example makes the role of the engineer explicit: the system proposes, tests, and refines variants, while a person reviews the final options and selects a design. The simulation result is not only a report; it feeds back into the design process for another iteration.

These demonstrations show why a general model is only one component of a useful industrial agent. The work depends on specialized domain data, tools, simulations, and human review. The agent can coordinate those pieces, but the systems and expertise that make the work possible remain essential.

Huang described physical AI as the next major extension of AI: systems that understand the physical world, reason about its complexity, and use appropriate tools to interact with it. Robotics is one example, but he also pointed to telecommunications. A radio network senses the world through radio waves and acts on it through beamforming, making it a physical system that can benefit from AI.

NVIDIA’s telecommunications proposal is to make cell towers part of distributed AI infrastructure. Huang said there are around 100,000 sites and roughly 100 gigawatts of distributed data-center capacity across the Western world. He argued that AI could improve spectrum efficiency and make the existing network useful for more than conventional communications. NVIDIA, Nokia, and T-Mobile, he said, made a first commercial phone call on the T-Mobile network after nearly eight years of work. The intended progression is to establish conventional calling on the platform, then add AI workloads.

The plan combines network infrastructure and accelerated computing at distributed sites. The slide described the initiative as “Telecom Networks Become Distributed AI Infrastructure” and named NVIDIA Aerial AI-RAN, CUDA-RAN, an Aerial RAN computer, and models and software including Nemotron, Cosmos, and TensorRT. Huang presented improved spectrum efficiency as a major opportunity for the partnership.

Healthcare provided a second example. Huang described adaptive radiotherapy as a robotic application because the tumor and the patient’s body change during treatment. The system must perceive the changing environment, locate and segment the tumor, and adjust the radiation precisely. NVIDIA and Siemens Healthineers announced work on a next-generation adaptive radiotherapy system, using components identified on the slide as Nemotron, Omniverse, Holoscan, Cosmos, Isaac Lab, and IGX Thor.

The analogy to robotics is about the system’s need to sense and respond to a changing physical environment. A treatment plan cannot simply be applied without regard to changes in the patient and tumor; the system must update its understanding and direct the treatment accordingly. Huang called it a form of robotic surgery, emphasizing the role of perception, segmentation, and precise action.

Autonomous vehicles are a more mature physical-AI effort in his account. Huang said the cost of building autonomous systems, especially the software, creates a case for standardization. NVIDIA’s Hyperion platform combines an autonomous-driving computer with a sensor configuration, safety systems, and software, then connects vehicles to ride-hailing services and software partners. He described a standard sensor configuration of 14 cameras, nine radar systems, three lidars, and ultrasonic sensors, supported by dual NVIDIA GPU systems.

The business premise is that robotaxis will join existing ride-hailing networks and operate across multiple manufacturers, rather than coming from one brand alone. Huang cited the world’s roughly trillion annual vehicle miles as an opportunity, arguing that standard hardware and software could let different automakers contribute to a shared fleet. Hyperion is intended to provide a common platform for that ecosystem.

He announced four additional vehicle brands for the platform: Volkswagen, JLR, Volvo, and Mercedes-Benz. The accompanying slide placed those automakers alongside ride-hailing services such as Bolt, Grab, Uber, and Tier Mobility, and autonomous-driving software companies including Applied Intuition, Pony.ai, Momenta, Nuro, AutoX, Wayve, DiDi, and WeRide. The point of the ecosystem model is that vehicle manufacturers, software developers, and fleet operators can connect to a common platform rather than each building every part independently.

Humanoid robotics, however, is the most ambitious manufacturing opportunity Huang highlighted. Robotics has been fragmented because different tasks often require different configurations, making it harder for volume to accumulate around a shared platform. Standardization, in his view, can create a technology flywheel: more customers support more production volume, and volume makes systems improve faster and become less expensive.

A humanoid form is not ideal for every task, Huang acknowledged, but could serve a broad enough range of work to become a high-volume configuration. If robots can learn by demonstration, small and midsize businesses without robotics teams could show them what to do rather than program every task from scratch. He pointed to Europe’s expertise in precision manufacturing and mechatronics as an advantage in building this industry.

Developing physical AI, he said, requires three computers: one to train the model, a digital twin in which the model learns and is tested, and a computer installed in the robot or vehicle to run it in the real world. NVIDIA’s Omniverse provides the digital-twin environment; the Isaac platform supplies tools for robotics development and deployment.

That separation matters because a model trained on data is not automatically ready to control a physical machine. The digital twin provides a place to develop and test skills, while the onboard computer must execute them in the real environment. Huang presented the reference platform as a way to avoid having each developer start from scratch with computing, sensors, software, and mechanical systems—a process he said could otherwise take several years.

Huang announced Agile ONE, a humanoid robot developed with Agile Robots, as a reference system intended to give developers a starting point rather than forcing each team to assemble hardware and software from scratch. The announced version, Agile ONE SN, was described as having a 31-degree-of-freedom body and five-finger hands with 20 degrees of freedom, and as available in the first quarter of 2027. The slide identified Isaac OS, Jetson Thor onboard compute, and Halos Safety OS as part of the system.

The reference robot sits within a broader platform strategy. Huang showed model builders and robot-platform companies alongside NVIDIA’s Isaac tools, and argued that standardized compute and software could make it easier for many companies to develop robots. The goal is not one robot for every application, but a common enough base to support wider development and production.

The RTX Spark announcement brought the agent architecture back to the individual computer. Huang said NVIDIA began thinking four years earlier about how to reimagine the PC for a world in which a computer is not only a tool but a personal assistant. Such a system needs a model, a harness, a secure runtime, and access to knowledge and tools. The model can run locally or in the cloud, but the computer must be able to execute useful tools.

NVIDIA’s pitch is that its history of accelerating graphics and compute workloads makes it possible to run a broad range of existing software on the new systems. Huang described RTX Spark as a new line of desktops, laptops, and workstations developed with Microsoft and the PC industry, powered by a chip called N1X. He said the system offers one petaflop and 128 gigabytes of memory, comparing it with the DGX-1 he delivered to OpenAI in 2016: the same stated performance and memory, but in a device he could hold.

He also described a Windows-based DGX workstation with 20 petaflops and 754 gigabytes of memory, aimed at users running applications such as CATIA, Autodesk, Adobe, Cadence, and Synopsys. The intended role is not just to answer questions locally, but to run agents and the accelerated tools they need on a personal machine, with the option to burst larger workloads to the cloud.

A demonstration used a small robot called Microduck. The presenter asked an agent to set up Hermes with OpenShell, use CUDA-, RTX-, and OpenGL-accelerated tools, and select an open model that would fit on RTX Spark. The workflow used Hermes as the harness, Nemotron and OpenAI’s GPT-6 Astra as models, and tools including FreeCAD, Blender, Isaac Sim, Newton, and Isaac Lab.

The system generated and configured a virtual environment, modeled objects, set up the robot’s joints and actuators, and simulated its movement. Training then ran in the cloud, where many Microduck instances learned in parallel in Isaac Lab. The resulting policy was downloaded to RTX Spark and evaluated in simulation before being deployed to the physical robot.

The demonstration made the three-computer model concrete: local computing for orchestration and tool use, cloud computing for large-scale training, and a digital twin for simulation and evaluation. An agent helped assemble a robotics workflow across those environments, while a physical robot carried out the trained behavior.