NVIDIA Releases DSX Pod and OpenShell at GTC Berlin; Jensen Huang Calls on All European Industries to Embrace AI

At the GTC conference in Berlin, NVIDIA CEO Jensen Huang rolled out a dense slate of new products spanning AI infrastructure, agent security, robotics, and autonomous driving, and issued a strong call to Europe's political and business leaders—every industry must become an AI company.

On October 21, NVIDIA CEO Jensen Huang said at the GTC conference that in the future all telecom operators will provide AI computing power, and telecom base stations are expected to evolve into distributed AI networks. He described the AI economy as a five-layer structure consisting of energy, computing power, AI factories, models, and applications, and in his presentation he highlighted that the computing layer features the logo of only one company: NVIDIA.

The announcements covered the modular AI data center product DSX Pod, the agent security architecture OpenShell, a humanoid robot reference platform, and new products and partnerships across verticals including autonomous driving, cancer treatment, and protein design. NVIDIA also announced at the event that over the past year it has invested nearly $10 billion in Europe and established more than 12,000 partnerships. The keynote was briefly interrupted by Greenpeace protesters; Huang said he would speak with them afterward and then continued his presentation.

Against the backdrop of growing doubts among some investors about the scale of AI infrastructure spending, Huang sent a clear bullish signal. "For Europe to benefit from AI and truly build an AI ecosystem and an AI economy, every end market and every end industry needs to adopt AI," he said during the keynote, naming the automotive, financial services, and healthcare industries. NVIDIA shares rose more than 2% in premarket trading at one point.

DSX POD: Compressing deployment from six months to 14 days

The DSX POD unveiled by NVIDIA is a complete AI factory reference design based on the Vera Rubin architecture. Its core logic is to move the integrated installation of computing, networking, cabling, rack-level power, and liquid cooling systems from on-site construction to the manufacturing plant. Partners complete inter-rack cable connections, electrical wiring, piping installation, and functional verification before shipping; once the container arrives on site, it only needs to be connected to pre-prepared power, cooling, and network facilities to go into operation.

Huang said:

"AI factories create value when they start producing tokens, not when construction is finished. With DSX POD, we bring factory-level precision and speed to AI infrastructure buildout, cutting deployment time from months to weeks so customers can scale capacity faster and bring revenue-generating production online."

It should be noted that the 14 days refers to on-site deployment time when the site has already been prepared, and does not include the full cycle from site selection to completion. Customers still need to complete preparatory work on supporting facilities such as power, cooling, and networking in advance. NVIDIA plans to deliver modular AI factories based on the DSX POD design globally in early 2027 with partners including Dell Technologies, HPE, Lenovo, Bull, Schneider Electric, and Vertiv.

On the energy efficiency front, the DSX AI factory platform uses closed-loop liquid cooling and dry coolers to reduce cooling water consumption to near zero, saving up to 54 million liters of water per day at 2 gigawatts of scale; DSX MaxLPS software can help operators reduce power consumption by up to 30% for the same token output.

OpenShell: Controlling AI agent behavior with hardware-level isolation

As agent deployments scale up, NVIDIA introduced the OpenShell reference architecture specifically to address agent security. The architecture integrates Vera Rubin infrastructure, Vera CPU, BlueField-4 data processors, and OpenShell secure runtime software, restricting the range of resources agents can access by enforcing policies at the granularity of system calls, file access, and network connections.

The core mechanism is the NVIDIA Sentry security framework running on BlueField-4, which continuously monitors agent behavior in a hardware-isolated trusted domain and, upon detecting anomalous operations, can isolate the agent and cut off its connection to the model within milliseconds. NVIDIA says the Vera Rubin NVL72 system can achieve up to 30 times the throughput per megawatt under agent workloads compared with before, while NVIDIA Sentry can increase agent capacity by up to 20%.

OpenShell is open-sourced under the Apache 2.0 license, supports enterprise environments such as Linux, Kubernetes, and Microsoft Windows, and can be extended to edge devices. Microsoft, Google Cloud, Dell Technologies, Cisco, OpenAI, Anthropic, and others have announced participation in related collaborations; among them, Anthropic and NVIDIA are working together to develop the Claude OpenShell agent.

Robotics and autonomous driving: multiple European automakers and robotics companies join the ecosystem

In the field of physical AI, NVIDIA disclosed that revenue in related markets grew about 30% year over year in the first half of this year, to about $11 billion.

On the robotics side, NVIDIA announced the expansion of the Isaac open robotics development platform, launching the open-source runtime framework Isaac OS, the robot learning framework Isaac Lab 3.0, and the SimReady Central Library containing more than 2,000 simulation-ready robots and factory assets. At the same time, NVIDIA partnered with Agile Robots to launch the humanoid robot reference platform Agile One SN, which includes Jetson Thor computing hardware and is expected to launch in the first quarter of next year. Huang said:

"General-purpose robots are advancing at an astonishing pace—the ChatGPT moment for robots is approaching fast."

In autonomous driving, Volvo Cars will use NVIDIA's Hyperion computing platform and the Alpamayo autonomous driving software stack to develop self-driving cars; Volkswagen will use NVIDIA products for commercial vehicles; and Bolt Technology is advancing its Robotaxi business based on NVIDIA software and hardware, with plans to begin collecting real-world driving data in Madrid starting in 2027 and to deploy at least 25,000 vehicles with L4 autonomous driving capability in Europe. In addition, Mercedes-Benz will use the Hyperion platform to develop L3 and L4 autonomous driving capabilities. NVIDIA also introduced the Halos SAFE safety evaluation framework, which integrates simulation, software- and hardware-in-the-loop testing, road testing, and fleet monitoring into a unified evaluation process.

Healthcare and biomedicine: from radiation therapy to protein design

In healthcare, NVIDIA announced several collaborations. With Siemens Healthineers' cancer treatment business, it is exploring the application of agentic AI, individualized patient simulation, and real-time computing to adaptive radiation therapy; the relevant systems will monitor changes in patients' breathing and position in real time and dynamically adjust radiation doses.

On the biomedicine side, NVIDIA partnered with the Institute for Protein Design at the University of Washington School of Medicine to launch the open model BioNeMo RFD4-Proteina, which can generate and optimize protein design proposals based on molecular inputs, design constraints, and scoring objectives, with applications covering drug discovery, disease detection, and industrial enzyme development.

For European enterprises and research institutions, NVIDIA announced an expansion of its open model portfolio, with new products covering the Nemotron 3.5 series of language and vision models, the Kumo model for structured business data analysis, the Cosmos world model for robot training and simulation, Alpamayo 2 Edge for autonomous driving edge inference, and the Ising model for quantum computing.

Jensen Huang likened the AI economy to a structure composed of five layers: energy, computing power, AI factories, models, and applications, and noted that AI infrastructure construction in Europe is accelerating. He predicted that in the future all telecom operators will provide AI computing power, and existing telecom base stations are expected to gradually evolve into distributed AI networks. At the partner level, AI data center builders such as Nscale and Firebird, as well as European AI model and application companies such as Mistral, Legora, and DeepL, were all included in the AI economy ecosystem map he presented.

The following is the full text of Jensen Huang's keynote at NVIDIA GTC Berlin 2026:

Jensen Huang:

In 2020, we began working with Mercedes-Benz to develop autonomous vehicles. Looking back on NVIDIA's 33-year history, I have spent a lot of time in Germany. So I am delighted to be here again, and delighted that GTC is being held here this year.

In the past year alone, we have invested nearly $10 billion in Europe, covering startups as well as large and small partnership projects across Europe.

Just in the past week, I traveled from France and the UK to Sweden, and then to Berlin. During that time, we announced investments in 14 startups. To date, we have established more than 12,000 partnerships in Europe, and everyone is working toward the same goal.

After decades of development in the computer industry, we are now standing at the starting point of a new industrial revolution.

Very interestingly, in AI's understanding, building a new industry means generating tokens, producing intelligence, and then integrating intelligence into every industry to drive a new industrial revolution.

A new industrial revolution occurs when a new technology changes how we make products and what products we are able to make.

I believe artificial intelligence will affect everything we make in every way, and will enable us to create new things that were previously impossible to create.

Everyone will use it, and every company will be empowered by it, because it is digital intelligence and the foundation of everything. Every industry will be changed by it, and new industries will be born.

It is not just a software technology. AI of course runs on computers, but the way these computers operate has fundamentally changed. It is no longer merely a tool for us to use; it has become part of machines, infrastructure, and factories, used to produce intelligence and then integrate that intelligence into every industry and application.

Energy powers advanced computers; these computers form AI factories; the models running in the factories generate tokens, that is, digital intelligence; and this intelligence is then integrated into various applications.

We are creating jobs and growth opportunities at an astonishing pace. Companies in Europe such as Schneider Electric, Siemens, and Siemens Energy are performing outstandingly around the world.

NVIDIA is at the computing layer, and our technology is being integrated into AI factories emerging across Europe. Emerging companies such as Nebius are doing very well; a new company called Firebird is building a new cloud platform in Central Asia, and in Africa there are companies such as Cassava.

I predict that in the future every telecom operator will transition to AI. The distributed infrastructure made up of a large number of communications sites and data centers will evolve into a distributed AI network. The original draft described the scale as about 100,000 sites and about 100 gigawatts.

We must recognize that infrastructure must exist before models appear. Just as networks had to be built before there was the internet, and power plants had to be built and homes connected to the grid before we could enjoy the capabilities of electricity.

The model layer is not just language models either; it also includes biology, chemistry, physics, robot motion, and animation models, and they vary in scale.

When it comes to NVIDIA's contributions at the foundational layer, most people first think of GPUs, and that is of course correct.

In 1993, we started with graphics chips, and later developed GPUs and programmable shader GPUs. Then we added CUDA acceleration, redefined computing, and drove the development of artificial intelligence.

Our technologies include GPU computing, with the latest generation being Vera Rubin; NVLink switches that connect multiple GPUs into one giant GPU; and the recently launched Vera and BlueField-4, which I will discuss later.

There is also our Ethernet solution Spectrum-X, which has made NVIDIA the world's largest networking company today.

There is also the Groq 3 decoding accelerator for text and token generation.

In addition, there is storage acceleration based on Vera and BlueField-4, as well as Vera RTX computing. I will introduce each of these separately later.

These systems are extremely large. A 1-gigawatt data center, that is, a 1-gigawatt AI factory, costs about $50 billion. This includes land, power, buildings, and computing, networking, storage, power supply, and cooling equipment.

Making this factory operate, be put into use, and produce tokens requires a massive amount of software; this is an extraordinary engineering effort. NVIDIA provides complete full-stack software.

On top of that are the libraries we call CUDA-X, which is NVIDIA's algorithm system.

In essence, NVIDIA is an algorithm company; we design chips to accelerate these algorithms.

One of them is cuDNN. Many researchers believe this library drove the modern AI explosion.

cuDNN is to AI what SQL is to the global database industry. It is a domain-specific language and library.

Just as OpenGL and DirectX drove the development of real-time computer graphics, cuDNN drove the development of the AI computing industry.

Beyond that, we have 350 other libraries. These libraries are the company's core assets, covering a wide range of areas such as computational lithography, signal processing, image processing, data processing, particles, fluids, and sparse solvers.

Many of your companies are already connected to these libraries.

In this way, we open up new markets, connecting NVIDIA with various algorithms so that everyone can complete computations at extremely high speed.

On the far right are the models.

Models are the new software and the engines of new applications. They are built on top of the CUDA-X libraries and work in concert with them.

NVIDIA has been developing these models in multiple fields, and they will open up new opportunities, which I will discuss in detail later.

Looking back at the generations of products in the AI era, we started with Pascal.

DGX-1, based on Pascal, was the world's first AI supercomputer. I launched it ten years ago and delivered the first one to OpenAI, starting their computing journey.

From Pascal to Volta, Ampere, then Hopper, which is what many people first encountered when they heard of NVIDIA, then the current Blackwell, and now Vera Rubin, which is already in full production.

This is six generations of architecture. With each generation, we have advanced technology.

Over the past decade, we have improved AI performance by 1 million times.

Think about it: if we had relied only on Moore's Law, what would have happened in ten years? A 10x improvement in the first five years and another 10x in the next five years, for a total of about 100x.

But over the same period, through what we call "extreme co-design," we innovated simultaneously across computer architecture and every layer of the software stack, optimizing algorithms, creating new algorithms, and designing corresponding architectures to accelerate them, ultimately improving AI performance by 1 million times. We will continue to do so.

The early Pascal, Volta, Ampere, and Hopper were mainly centered on GPUs. We created a new DGX form factor.

With Blackwell, we introduced the world's first rack-scale computing system.

Rack-scale computing turns an entire system into one giant computer, or one giant GPU.

With Vera Rubin, we further expand the scope of computing: no longer just handling models or large language models, but handling the entire agent.

We will come back to this diagram later, but remember: agents need to think, and models do that part of the work.

Perception, reasoning, planning, and tool use all depend on large language models, and models are designed for these capabilities.

But an agent also needs a runtime framework, a Harness. You can think of it as the exoskeleton of a digital agent.

Just as humans have a brain and a body, the agent framework coordinates the work, enabling the agent to call on memory, including active memory, working memory, and long-term memory, and to use tools.

The third part is the computer that processes knowledge, specifically responsible for data processing.

Finally, there is a class of computers dedicated to running CUDA-X libraries, responsible for executing tools.

Together, these computers form a complete agent system.

Vera Rubin is already in full production. Let's take a look at the real thing.

This is the first computer we have built with no exposed cables and a 100% liquid-cooled design. It can be pulled straight out, which is very convenient.

This is a Vera Rubin computer; this unit has two GPUs.

In the past, assembling a Grace Blackwell node took two hours. Once you've seen one in person, you'll understand—cables everywhere, and it uses an approximately 80% liquid-cooled design.

Now, Vera Rubin is 100% liquid-cooled, and assembly takes less than five minutes.

This is remarkable progress, meaning shorter manufacturing cycles, higher reliability, and better performance.

This is the NVLink switch tray. It connects 72 GPUs to form one giant GPU.

Each GPU delivers 20 PFLOPS of performance, and then 72 of them are connected into one massive system.

This is our new Vera computer.

[Interjection from the audience]

It's all right, be gentle, give her a chance to speak. Everyone deserves a chance to express themselves. But today, let me finish my presentation first, and I'll talk with you afterward.

These are all important questions, but please let me finish the keynote first.

This venue is already full, and there are 100 synchronized viewing events around the world, with 15,000 people watching right now.

Finally, I've prepared some very interesting content. Before the entertainment segment arrives, everyone will have to endure a bit of my technical explanation.

This is the all-new Vera computer. The Vera CPU is an extremely powerful processor, capable of executing eight instructions per clock cycle.

It uses what we call the world's first LPDDR memory subsystem, combining high performance, extremely high bandwidth, and outstanding energy efficiency.

Its on-chip interconnect speed leads the world among chips, so data processing, as well as the startup and execution performance of agent sessions, are all outstanding. That is exactly what it was designed for.

This is the Vera computer.

Next to it is Spectrum-X, which uses co-packaged optics, or CPO technology. Let me show you.

Some spots on the stage are empty, and some can be stepped on—I'm not quite sure which is which. If I suddenly disappear, you should know it's not a stunt, it's an accident.

Here are the lasers. Optical signals go directly into the chip through optical connections.

We developed a process called COUPE together with TSMC, allowing silicon photonic devices to connect directly to the chip.

Otherwise, this entire area would need to transmit signals through printed circuit boards, and circuit boards are far less capable of transmission than optical fiber.

Now, we can deliver extremely high data rates and high-quality signals directly to the chip, reducing the conversion steps through circuit boards and transceivers in traditional approaches.

This is our CPO product, already in full mass production, offering 512 ports at 200Gbps per port.

This is the world's most advanced Ethernet switch, and we are very proud of it.

These systems allow us to handle the thinking process of large language models, run agent frameworks, process the knowledge data agents need, and run various tools through accelerated computing. These basically make up the complete workload of an agent.

Vera Rubin is already in full production, with outstanding performance.

As I said, with each generation, we achieve performance gains far beyond what Moore's Law could deliver through extreme co-design.

Vera Rubin takes another big step beyond Blackwell.

In large language model processing, Blackwell delivers about 30 times the performance of Hopper. And just two years later, Vera Rubin reaches about 30 times that of Blackwell.

Next, let's look at how Vera Rubin forms a DSX full-stack AI factory.

Video narration:

NVIDIA DSX is a full-stack AI factory reference architecture developed together with the world's leading infrastructure companies.

Compute, networking, storage, software, power, and cooling are co-designed as a whole, optimized end to end from the design phase through actual operations.

DSX starts with a digital twin.

Before construction, performance, power, water usage, and scalability are validated in a virtual environment.

PTC turns equipment designs into assets that can be used for simulation.

Dassault Systèmes connects rack design with facility design through model-based systems engineering.

Schneider Electric simulates power distribution and liquid cooling systems.

Jacobs integrates these into an Omniverse digital twin.

DSX Max LPS is an end-to-end control loop that dynamically allocates power based on workload demand, reducing electricity use by 30% while achieving the same output.

DSX also redesigns cooling.

NVIDIA Vera Rubin systems use closed-loop liquid cooling, and in suitable climates, the coolant temperature can be as high as 45 degrees Celsius. Outdoor dry coolers release heat into the air without evaporating water.

For a 2-gigawatt facility, DSX can save 54 million liters of water per day, equivalent to the daily water use of 180,000 households.

However, building an AI factory can take months. Until it is commissioned, it produces no tokens.

Today, we are launching NVIDIA DSX Pod, which integrates this full-stack platform into an all-in-one system manufactured in the factory.

It is built on Vera Rubin NVL72.

Seven liquid-cooled computer racks, connected by 17 kilometers of Spectrum-X cabling, form a Vera Rubin DSX Pod.

The entire Pod is integrated, cabled, and tested at the factory, then delivered in a transportable container.

Dry coolers replace traditional evaporative cooling systems, and 800-volt DC power connects directly to the racks, improving power efficiency.

As long as the on-site infrastructure is ready, on-site deployment time is expected to be shortened from months to days.

The NVIDIA DSX AI factory takes resource efficiency and responsibility into account from the design stage. Now with DSX Pod, AI factories will be built much faster: more intelligence per watt of power, and less time from delivery to generating the first token.

Jensen Huang:

So, what does everyone think?

In 2016, we released the world's first AI supercomputer. At the time, it was a single-node GPU system.

Three years ago, we introduced the Grace Blackwell NVLink 72 rack-scale computer.

With these two innovations, we not only redesigned the computer, the software, and how it operates, but also redesigned the supply chain.

Take NVLink 72 as an example. In the past, the entire rack had to be assembled on site; now we have moved the assembly work into the supply chain.

In essence, we have achieved large-scale manufacturing of supercomputers.

This is exactly why we can move so fast while supporting hundreds of data center construction projects around the world.

Today, all major computer manufacturers worldwide are building NVLink 72. We are producing computing systems at a scale that was previously unimaginable.

Even so, once these systems arrive on site, it still takes months to assemble them.

As you just heard, the Spectrum-X connectivity for scale-out computing alone requires 17 kilometers of cabling.

Suppose we move all of this work into the supply chain: assembling the system, powering it on, verifying full functionality, completing quality certification and burn-in testing, and ensuring it works properly when shipped.

This would greatly reduce the amount of on-site cabling and assembly.

But more importantly, it would significantly reduce the idle compute capacity in data centers that is waiting to be installed and commissioned and cannot be put to use.

Remember, each 1-gigawatt data center costs about $50 billion and is expected to generate tens of billions of dollars in revenue per year.

Suppose this infrastructure could generate $40 billion in revenue per year; a delay of just three months would mean $10 billion in revenue is postponed. That is enough to make anyone anxious.

By moving this work into the supply chain, the time from computing equipment arrival and installation to officially producing tokens is expected to be shortened from four to five months to about one month.

This will fundamentally change the global supply chain, and also change the way people install and build data centers.

We hope this can increase the speed of construction and commissioning, reduce the waste of idle compute capacity and idle capital expenditure, and give all participants better profitability.

This is DSX Pod, which further integrates NVLink 72 into a large cabinet system, and it is the next step in the evolution of NVIDIA computing systems.

What you make, and how you make it, determines what makes you unique.

Like European companies, the expertise you possess sets you apart: you design machines, operate factories, discover drugs, and possess extraordinary craftsmanship and expertise.

NVIDIA's uniqueness lies mainly in two points.

First, our architecture is a general-purpose accelerated computing architecture.

We do not target just one algorithm, such as Transformer, but support a wide variety of workloads that can be processed in parallel, including fluids, particles, quantum computing, quantum chemistry, image processing, data processing, and structured data and vector data processing.

Because of our long-term investment and because the architecture has remained consistent for about the past 25 years, the number of algorithms, domains, and industries we support and accelerate continues to grow.

A massive installed base, a general-purpose architecture, long-term investment, and collaboration with various industries make the NVIDIA architecture almost ubiquitous.

Second is how we deliver it.

Designing chips, manufacturing supercomputers, and writing all the software are not tasks most companies intend to take on.

Therefore, the combination of CUDA and the full-stack AI factory enables us to serve markets that were previously difficult to reach. This is also why we can work with so many companies here.

Starting with cloud service providers, or CSPs, most people believe that the vast majority of AI today is concentrated there. But at the beginning of the keynote, many of the companies I mentioned are not traditional CSPs.

CSPs remain an important part of the global computing market, but at the same time, AI-native companies are emerging around the world.

For example, France's Mistral, Europe's Nebius, and companies building new AI clouds across Europe.

There are also OEMs and ODMs that manufacture computers for enterprises worldwide, as well as AI labs.

Last year, we talked a lot about OpenAI; today, in addition to OpenAI and Anthropic, there are also players such as Meta, Grok, and Gemini, and more and more AI labs are driving frontier technology forward.

With CUDA and the way we manufacture computers, we can serve every segment of AI.

One architecture, covering multiple markets.

These infrastructures are enormous and expensive, so they must reach maximum productivity as quickly as possible and be flexibly usable for a wide variety of workloads.

That is why the NVIDIA architecture is highly general-purpose and adaptable.

Continuous innovation, continuous software updates, and the rich software ecosystem formed by a massive installed base give this architecture long-term value.

One important moment was earlier this year when we talked about OpenClaw, which brought the concept of open-source agents to more people and also drove the changes now happening around the world.

Agents think, reason, iterate, and use tools, so they consume a large number of tokens. But that is precisely why they are starting to become truly useful.

More agents are being developed, and because they are more useful, more people use them; and because thinking requires a large number of tokens, token consumption is growing rapidly, and demand for GPUs is rising sharply along with it.

This diagram shows the five components of an agent.

The first is the language model. The model runs on GPUs and generates the tokens that agents use.

The second is what I call the "exoskeleton," that is, the agent framework, such as OpenClaw.

The framework runs inside a sandbox, which isolates it from the rest of the system. This is a very important concept.

The other two components are knowledge and tools.

The model, framework, sandbox, knowledge, and tools make up the five elements of an agent.

Before discussing how to make agents useful, we must first ensure that they are safe.

Agents are autonomous systems that can access files, resources, networks, and tools. Therefore, we must grant them the minimum necessary permissions and continuously monitor their behavior.

To this end, we have launched a new platform: OpenShell.

OpenShell is a secure runtime that provides the capabilities just mentioned and can work with additional hardware monitoring mechanisms.

To be effective, they need to reason, plan, use tools, access data, and collaborate with one another.

But today, the way to stop agents from abusing these capabilities is often just to tell them:

"Please don't delete anything." "Please don't push code to the main branch." "Please don't touch my SSH keys."

This is the "please don't" approach to security.

But hope is not security. Security is not a wish; it is technology and engineering.

NVIDIA OpenShell and Sentry put this concept into real systems.

You write a policy for each agent that specifies what it can touch. OpenShell enforces these policies at the level of system calls, file access, and network connections.

Formal verification checks every policy change to ensure that it does not exceed the permission ceiling you have set, so that agents cannot gain access beyond what is expected.

This is similar to the early internet: viruses spread everywhere before browsers strictly limited what web pages could do. OpenShell establishes a similar boundary for agents.

These restrictions are enforced outside the agent rather than relying only on guardrails inside the large language model, thereby bringing native agent security capabilities into the infrastructure.

Each agent runs in its own isolated sandbox and cannot directly access the network.

All outbound network requests must pass through a supervisor component outside the sandbox. The supervisor checks requests against policy, and operations that are not permitted are denied.

Keys are kept outside the sandbox and managed by the supervisor, and are added only when requests are sent to endpoints that are permitted to access them. Even if an agent is compromised, it cannot leak keys it never held.

The gateway acts as the control plane, the supervisor makes decisions, the sandbox contains the agent, and the policy prover checks whether proposed policies comply with the permission ceiling.

OpenShell combines Linux kernel isolation with runtime policy enforcement.

NVIDIA Sentry can also add an optional layer of protection outside the host and away from the agent.

Sentry runs on NVIDIA BlueField-4, monitoring agent activity and enforcing security policies within a hardware-isolated trust domain.

Agent inference traffic passes through Sentry, so once a plan that needs to be blocked is detected, the system can immediately update policy, cut off the agent's connection to the model, and complete isolation within milliseconds.

It also keeps verifiable, tamper-proof records for operators to detect, investigate, and respond.

OpenShell is open-sourced under the Apache 2.0 license and can run in the various environments where agents reside, from the edge to the cloud. You can get started with just two commands: install the command-line tool, create a sandbox, and then run your favorite agent.

How much an agent can accomplish depends on how much access you give it. OpenShell and Sentry keep that access under your control.

Jensen Huang:

Agents are autonomous systems. We must create highly secure technology to ensure that they can be developed and deployed safely.

After OpenShell was released, many large enterprises began adopting it, from frontier model labs to operating system companies, cybersecurity companies, and platform companies.

Because this technology arrived at just the right time: we are preparing to deploy agents at scale around the world.

This is not only about service providers using OpenShell when deploying agents, but also about enterprises around the world adopting OpenShell so that these agents can operate autonomously within their businesses.

Different capabilities, different tasks, and different users correspond to different permission requirements. Each time an agent is launched, the permissions it needs may be different.

Therefore, permission configuration, policy enforcement, and monitoring must all be done dynamically.

I am delighted to announce that Anthropic and NVIDIA are collaborating to combine Claude agents with OpenShell. There is a great introductory article on the website.

You can use OpenClaw, Hermes, OpenHands, or build your own framework with LangChain, Deep Agents, or even write it directly in Python.

All of these agent frameworks can run in the secure sandbox provided by OpenShell.

Next, let's talk about models.

In the future, we will not use just one model for all work; instead, we will choose the most suitable model for the task, whether it is closed-source, open, or a custom model.

Some work is relatively general, and general-purpose AI can do it well.

But other work involves a great deal of proprietary knowledge and specialized capability, and you want to keep improving those capabilities.

Therefore, in the future every company will use both closed-source models and its own custom open models.

First, of course, are the models.

Nemotron is for language.

Cosmos is our world foundation model, and we are at the forefront of this field.

In these different areas, we offer models and continue to invest in research and development. We have been doing this for ten years, and we will keep doing it for the long term.

We not only open our models, but also provide the associated data so you can understand where the model's data comes from, and provide the scripts needed to reproduce it.

So what you get is not just open weights, but a complete methodology, data, and reproduction path. You can know where the model comes from and build trust in it.

These models are not necessarily at the frontier in every area, especially when it comes to language models.

But through customization and collaboration with ecosystem partners, you can raise them to the frontier level in the areas you need, turning proprietary data into intelligence you control.

For example, AI21 helped Alice customize Nemotron into a model that surpasses the relevant frontier capabilities.

Baseten helped CodeRabbit fine-tune Nemotron with its own data to achieve frontier-level capabilities.

Trajectory and Harvey also built their own frontier capabilities by post-training Nemotron.

Scale and Skydance are doing the same thing.

There are many more examples.

The key point is that you can start from an excellent open model like Nemotron and post-train it with your own data for a specific use case, thereby obtaining highly specialized intelligence.

This is a very sensible path, and all the required technology is available today.

There is also the case of Palantir and Tampa General Hospital.

We have a rich partner ecosystem that can help everyone. These are all outstanding companies that can serve customers around the world.

If I mention only one, I will leave out the others; if I list them all, it will take a long time. I deeply value all of these partners.

To summarize: first, run agent frameworks in the OpenShell secure sandbox; second, choose a large language model such as Nemotron.

But what truly turns an ordinary agent into a "super agent" is giving it access to your company's knowledge.

We call this capability the agent knowledge foundation.

It consists of a set of world-class models and tools that help you process your company's various data assets.

Data may exist in unstructured form in documents, so AI is needed to understand PDFs and other unstructured files; it may also exist in structured form in databases.

We provide the necessary libraries and tools to help you extract meaning and build understanding from enterprise data.

Nemotron Parse handles document parsing and ranks at the top of the relevant leaderboards.

Nemotron Embed links queries with knowledge in vector databases and also delivers leading performance.

Nemotron Rerank processes multiple retrieved sources and data fragments, reordering them according to their relevance to the query.

Kumo can use structured databases to predict what may happen.

For example, you receive a supplier notice that certain materials will be in short supply or have delivery delays. This information is likely an unstructured document.

Company order information is usually stored in structured databases such as SQL.

If you want to predict which orders will be affected, which will be out of stock, and for how long they will be delayed, you can use Kumo.

Next, to find alternative parts, you can use vector database retrieval to find engineering documents for relevant replacement components.

Finally, there are tools, the tools needed to complete specialized work.

NVIDIA has been developing CUDA-X.

These tools are used for discovery in various scientific fields, as well as for engineering, research, design, construction, and the creative arts.

At their core are the CUDA-X libraries, and CUDA-X libraries are essentially algorithms. As you know, algorithms are beautiful.

Let's look at a demonstration.

Mathematics is beautiful.

The CUDA-X libraries are our company's core asset. The reason is that when you build an AI assistant—whether it's a lab assistant, a research assistant, or a partner that designs and creates with you—what ultimately makes it truly useful and able to complete real work is tools.

Now, we have models that enable agents to understand information, reason, plan, and continuously take actions and use tools until the task is complete.

But in the end, it still comes down to tools.

These tools are something we have been developing for many years.

The kind of computer shown on the right just now can run all these tools, whether it's fluid dynamics, particle physics, quantum chemistry, data processing, or ray tracing.

Compared with running directly on a CPU, it can increase processing speed by 10x, 50x, or even 100x.

CUDA-X gives agents the ability to complete specialized work. Seeing how others use these tools is very exciting.

This is the key behind agents.

Agents ultimately need to use tools.

Software programming is one of the earliest fields to adopt agents, where a compiler or Python is often enough to accomplish a lot of work.

But in the future, as agents take on more and more real tasks, we need to provide them with more tools.

Here are a few examples.

Lila Sciences is developing a lab assistant that uses NVIDIA's BioNeMo models to understand proteins and cells, and to design proteins and chemical molecules in drug discovery, while calling CUDA-X tools.

Siemens is building factory agents that use the Cosmos world foundation model, which understands the physical world, along with CUDA-X tools.

Dassault Systèmes is developing a virtual companion that uses the Nemotron language model and CUDA-X tools.

Let's see how they work.

Video narration:

The knowledge base provides context from the company's structured and unstructured data.

Tools enable agents to simulate, experiment, and take action.

NVIDIA OpenShell controls what agents can access and execute.

You can start with frontier closed-source models or customize open models. With the NVIDIA Agent Toolkit, any company can turn its own knowledge and expertise into a custom super agent.

Let's look at three examples.

For Lila Sciences, the goal is to design a therapy for B-cell lymphoma.

The Lila Iris agent, built on Nemotron, turns this goal into an experimental plan.

In the computational dry lab, the Lila agent uses the NVIDIA BioNeMo Agent Toolkit to call protein design and RNA optimization models, running millions of virtual experiments.

The most promising candidates enter Lila's AI science factory, where they are prepared and tested by autonomous robots and instruments developed with Omniverse and Cosmos.

The results return to Iris, informing the next round of design and experimentation, forming a continuous iterative closed loop.

At Siemens' Erlangen electronics factory, the goal is to bring new products into production.

The Process Preparation X agent uses Cosmos to generate virtual prototypes and synthetic data, training and validating inspection models before production begins.

The Intelligence Center X agent, built on the NVIDIA factory operations blueprint, proposes production plans based on orders, material supply, and line configuration.

Digital Twin Composer uses the Omniverse libraries to combine physical simulation with real-time factory data, validating these plans before actually adjusting the production floor.

On the Dassault Systèmes platform, an engineer needs to adapt the door design of a Volkswagen Golf to the Tiguan model.

Aura and Leo use Nemotron to understand and reason about engineering requirements.

The agent framework coordinates specialized sub-agents and preserves context throughout planning, design, and simulation.

Aura uses ENOVIA to plan the work based on vehicle dimensions, weight limits, and other engineering requirements.

The Leo design agent selects a compatible Golf door design, then uses CATIA to modify the shape so that it fits the Tiguan.

The Leo simulation agent uses CUDA-X-accelerated SIMULIA to test different options and discovers a vibration issue.

The results return to the design agent for optimization, and the engineer reviews the improved design and selects the final option.

Different goals, different expertise, the same foundational components: open models, CUDA-X-accelerated tools, and the OpenShell secure runtime, integrated through the NVIDIA Agent Toolkit.

Jensen Huang:

Pretty great, right?

It's actually the opposite.

Cadence, CrowdStrike, Dassault Systèmes, and many other enterprise software companies are building agents. We are delighted to work with them. [Another company name in the original list is unclear.]

Building agents is not that difficult. It has five components: models, frameworks, the OpenShell secure runtime, and most importantly, knowledge and tools.

Models and frameworks come in both open and closed-source versions, while knowledge and tools belong to you and represent your expertise.

But one of my favorite applications is actually the mobile communications base station.

It doesn't use cameras, speakers, or microphones; instead, it senses the world through radio waves, processes information and performs reasoning, and then interacts with the world using beamforming.

We will bring massive spectral efficiency gains to telecom networks.

Beyond that, it will also transform these communications sites. The original draft states that there are about 100,000 sites in Western regions, constituting roughly 100 gigawatts of distributed data center infrastructure, which will evolve into distributed AI infrastructure in the future.

We are delighted to work with many companies in Europe.

After nearly eight years of R&D, just last week, NVIDIA partnered with Nokia to complete the first commercial call on the T-Mobile network.

Now, CUDA is already able to perform traditional communications processing. On that basis, we can further layer on AI capabilities to drive new transformation.

During treatment, the tumor shape may change, and the human body also moves. This adaptive radiotherapy system must sense environmental changes, identify the tumor location, precisely segment the target, and accurately guide radiation therapy accordingly.

In essence, this is a robotic form of treatment.

We realized that every car company wants its vehicles to join mobility service fleets.

Today there are already platforms such as Uber, Bolt, Grab, and Lyft. In the future, robotaxis will also connect to these mobility services, and they will not come from just one brand.

But developing such a system, especially its software, is very costly.

So we decided to create a standard architecture.

This architecture includes 14 cameras, 9 radar sets, 3 lidars, and a set of ultrasonic sensors.

These sensors form a perception system surrounding the vehicle, with autonomous driving compute provided by a dual NVIDIA GPU system.

We have developed it into an industry standard, and it is now being adopted by more and more vehicle models.

Autonomous driving software companies develop capabilities on it, and NVIDIA itself is one of them. These robotaxis then connect to mobility service platforms, forming a complete ecosystem.

Global annual travel mileage is about 1 trillion miles. In the future, a considerable portion of that mileage may be completed by robotaxis, and at $1 to $2 per mile, this is a huge opportunity.

But the market is very fragmented.

By establishing a standard architecture like Hyperion and working with the entire ecosystem, we can form a huge virtual fleet that everyone can connect to.

Today, I am pleased to announce that four more brands or model lines are joining: Volkswagen vans, new models from Jaguar Land Rover, Volvo fleets, and the next-generation models from Mercedes-Benz.

They will be compatible with Hyperion and become part of this autonomous mobility network.

Welcome to Hyperion.

The next great industry, and very likely an extremely large one, is humanoid robots.

One important reason robots are hard to build is that there are so many kinds of tasks.

If a special form factor were developed for every task, the robotics industry would be split into many different configurations, making it difficult to achieve the economies of scale that drive technological progress.

Phones, PCs, and NVIDIA NVLink 72 have advanced so rapidly because standardized configurations can serve a large number of customers.

Large demand brings scale, and scale drives technological iteration, making products faster and better.

Therefore, we need robot form factors that are as standardized as possible.

Humanoid robots may not be suitable for all tasks, but they are likely suitable for many. Therefore, they may become the robot form factor with the largest shipment volume.

With general-purpose capabilities and general intelligence, robots can quickly learn and adapt.

This is very likely to become the next major technology platform.

We are working with many companies, quite a few of them from Europe. This is an important opportunity for Europe.

Europe has deep expertise in mechatronics, electronics, and the precision and craftsmanship of mechanical systems. There are many companies here with extremely strong manufacturing capabilities.

Robotics will become the next large-scale industry. Many companies have already joined in, and we are delighted to work with everyone.

We created the Omniverse digital twin platform.

Many of the CUDA-X demos we just saw run in a digital twin world.

Omniverse is a set of libraries that can connect to various platforms and tools, including platforms from Siemens, Dassault Systèmes, Cadence, and others. Together they build digital twin systems.

The third category is the computer deployed on the robot.

For autonomous vehicles and robotaxis, we call the relevant platform Hyperion.

We also need to provide an equally complete platform for robots.

That way, when everyone starts researching or developing robots, they won't have to assemble all the software, compute systems, sensors, and mechanical components from scratch. Just completing that preparation work could take several years.

So we decided to build a reference robot.

This is Agile One, developed by NVIDIA in collaboration with Agile Robots.

This reference robot is expected to be available in the first quarter of next year, already assembled at the factory, with a built-in computer and integrated software, and it will have quite good capabilities right out of the box.

In 1993, when we were founded, the PC revolution was about to enter a new phase; Windows 95 subsequently accelerated it further.

Since then, we invented the GPU, drove the development of programmable shading and real-time ray tracing, added CUDA to the computing model, and helped ignite the modern AI revolution.

Four years ago, we started thinking: in the face of a new era, can we redefine the personal computer?

It is no longer just your tool, but your personal intelligent agent, your personal assistant. Your interaction with it is more like a conversation with a partner who helps you work.

Therefore, this computer needs to have the five parts introduced today.

It needs models, and the models can run locally or in the cloud.

As long as we create a computer that supports OpenGL, DirectX, and CUDA acceleration at the same time, everything NVIDIA has developed in the past will be able to run on it.

This is the result.

After four years of collaboration with Microsoft, we have built RTX Spark, an entirely new computer product line.

Thousands of engineering person-years of investment are condensed into it. We developed it in secret for four years, creating a chip called N1X, which was recently announced together with the entire PC industry.

We have built desktops, laptops, and workstations.

This device delivers 1 PFLOPS of computing power.

For comparison: in 2016, I delivered a computer with 1 PFLOPS of computing power and 128GB of memory to what was then a still-small OpenAI.

Now, this small device also delivers 1 PFLOPS of computing power and 128GB of memory.

The difference is that I can hold it in my hand like this, whereas the DGX-1 back then required two people to carry it upstairs.

This is the all-new RTX Spark PC.

If you want a companion workstation for running CATIA, other Dassault Systèmes software, and tools from Autodesk, Adobe, Cadence, and Synopsys, you can also choose the DGX Station running Windows.

It delivers 20 PFLOPS of computing power and 754GB of memory.

In addition, there is a desktop version.

The related products are being rolled out, and the DGX Station running Windows is expected to go on sale by the end of this year.

Now, I have covered the five components of an agent. The next question is, how do you build an agent?

How do you get it to run on these computers, or in the cloud, or in enterprise on-premises environments?

You are about to see a robot application demo. As you watch, keep in mind that an agent is running on this personal computer, and the entire demo involves three types of computers:

Let's watch the demo.

Video narration:

This is Micro Duck, a robot from Hugging Face.

Let's use NVIDIA RTX Spark to teach it to observe the world and walk.

I use Hermes as the agent framework, with Nemotron and OpenAI's GPT-6 Astra as the models, and CUDA-X accelerated tools: FreeCAD, Blender, Isaac Sim, the Newton physics engine, and Isaac Lab.

Unless a large workload needs to be scaled to the cloud, all tasks run on RTX Spark.

First, I describe the training site and what Micro Duck needs to learn.

Hermes and Nemotron on RTX Spark plan the site and list all the required objects.

Then, Astra uses FreeCAD to build 3D object models, uses Blender to apply materials, and then adds physical properties through the Omniverse simulation-ready library.

Astra sets up the scene in Isaac Sim, configures joints, actuators, and connection structures for Micro Duck, and then uses the Newton physics engine to simulate its motion.

Next, training begins.

Isaac Lab in the cloud lets thousands of Micro Ducks train simultaneously.

Every fall is a learning opportunity, driving continuous improvement of the control policy.

The Omniverse library simulates Micro Duck's camera and teaches it to "see" the environment.

Once perception and physical capabilities are combined, Micro Duck can avoid obstacles, pass through narrow spaces, and adapt to its surroundings.

I download the trained policy to RTX Spark and use Nemotron and Isaac Sim to evaluate Micro Duck through multi-agent orchestration.

Finally, the validated policy is deployed to the real robot, and it can start moving.

NVIDIA RTX Spark lets you build and run AI agents and their accelerated tools directly on a laptop.

You were born from RTX Spark, trained on a large number of NVIDIA supercomputers in the cloud, and learned skills in Omniverse digital twins on RTX Spark.

Isn't that amazing?

You are so cute. Your "parents" are called Pollen, right? Pollen Robotics.

I heard everyone likes you so much that you are already sold out.

Everyone watching this video—there may be twenty million of them—can go to Pollen's website and place the next order.

This startup probably has only about ten people, and orders could overwhelm them all at once.

Everyone should have one; you are so cute.

With an RTX Spark, we can continue teaching you new skills.

We have released the new DSX Pod system to make building AI factories easier.

Over 33 years, this has been an incredible journey. We have made a cool video—do you want to watch it together?

Thank you, everyone; thank you to all the sponsors, and thank you to the partners who have worked with us along the way over these 33 years.

And Micro Duck, my new friends.

Have a wonderful GTC, everyone! Risk Disclosure and Disclaimer

Markets carry risk, and investment requires caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their particular circumstances. Investment based on this is at the user's own risk.