This week, there were two important events – a keynote at CES and a Q&A with CES financial analysts – where Jensen Huang shared his vision for the future. It’s clearly a union of Agentic AI and robotics, or as he calls it, Physical AI. In today’s episode of our Agentic series, I want to focus on Huang’s perspective on Agentic AI, the progression toward physical intelligence, and the actionable insights we should keep in mind as this shift unfolds. He predicts that within 10 years, robots might reach a level of capability that surprises even skeptics. And as Jensen Huang sees it, it’s a multi-million industry. Once again, I highly recommend watching his entire presentation (or should I say, show?) from CES – you’ll be looking into the future. Ready? Let’s peek together.
❝
“I worked backwards, and I said, ‘Okay, what does Nvidia have to do to realize that future for the world? What do I have to do to make that possible for the world to do?”
Jensen Huang
Jensen Huang, NVIDIA’s charismatic and visionary CEO, who is not afraid to be a little goofy at his keynotes, has long been at the forefront of defining the cutting edge in AI. In his most recent talk at CES, Huang articulated a transformative vision for the evolution of AI – a world where digital agents seamlessly execute complex tasks, and physical AI systems fundamentally reshape our interactions with the real world.
That’s how his roadmap for AI’s evolution looks like:
Perception AI: The foundation of modern machine learning, where systems are trained to interpret and analyze data.
Generative AI: Today’s focus – models creating text, images, and other content that are transforming industries like gaming, marketing, and media.
Agentic AI: The rise of autonomous agents capable of managing workflows, solving problems, and delivering insights.
Physical AI: The next step – AI systems gaining embodiment to operate and interact in the physical world.
Huang thinks human robotics is a multi-billion-dollar industry, but it’s impossible without working agentic systems. So, let’s take a closer look what gets us to it.
Since CES 2025, many of Huang’s predictions have started to play out faster than expected. Agentic AI has become a major focus for frontier labs, while robotics, world models, embodied AI, and physical reasoning have all moved much closer to the center of the industry. NVIDIA is trying to own the entire path from a model deciding what to do to a machine actually doing it: reasoning, agent orchestration, simulation, and physical execution.
Agentic AI vs Physical AI: What's the Difference?
Agentic AI refers to autonomous or semi-autonomous systems that can reason, plan, use tools, and complete multi-step tasks with limited human supervision. In Jensen Huang’s roadmap, it is the digital foundation for physical AI: robots and autonomous machines that can perceive, reason, and act in the real world.
So many people in their predictions proclaimed 2025 The Year of Agents. Huang is fascinated with it as well, describing these systems as a “new digital workforce” with the ability to reason about missions, breaking them into actionable subtasks, retrieve relevant data, and using tools to generate high-quality outcomes. Unlike traditional software systems, agentic AI adapts to dynamic contexts and operates autonomously or semi-autonomously, augmenting human capabilities across industries.
What Makes Agentic AI Different from Traditional AI?
Reasoning and Adaptability: These systems can decompose complex problems, analyze contextual factors, and prioritize tasks dynamically. This reasoning capability depends on how agents structure and apply their knowledge — a topic we explore in depth in Knowledge in AI Agents: Representation, Types & History.
Multimodal Capabilities: Leveraging foundational models, they integrate language, vision, and audio to process information from multiple sources and formats.
Tool and Data Integration: Agentic AI doesn’t just process static inputs; it interacts with tools, retrieves information from databases or the web, and synthesizes insights.
Real-World Applications of Agentic AI: Huang's Examples
Huang envisions Agentic AI as ubiquitous, spanning industries and use cases. Examples include:
Knowledge Workers: AI research assistants simplify the analysis of complex documents like academic journals or financial reports, turning them into accessible formats like summaries or podcasts.
Industrial Optimization: Agents monitor processes in manufacturing or logistics, identifying inefficiencies and recommending improvements.
Software Security: AI agents continuously scan codebases for vulnerabilities, providing real-time feedback to developers.
Healthcare and Drug Discovery: Virtual lab agents screen billions of compounds to identify promising drug candidates faster and more cost-effectively than ever before.
These examples highlight how agentic AI can streamline and improve tasks that have typically relied on human effort.
What Is Physical AI? Jensen Huang's Definition
While agentic AI focuses on digital tasks, the next frontier lies in physical AI. Huang describes physical AI as systems capable of understanding and interacting with the real world. This shift involves developing AI models that comprehend physical dynamics, spatial relationships, and environmental nuances – enabling them to operate in unstructured, real-world environments.
Physical AI builds on the foundation of agentic systems but incorporates an understanding of the physical world. It requires models to grasp:
Geometry and Spatial Reasoning: Understanding three-dimensional spaces and how objects interact within them.
Physical Dynamics: Concepts like gravity, friction, and inertia that govern real-world movement and behavior.
Temporal Awareness: The ability to predict and adapt to changes over time, such as shifting weather conditions or object trajectories.
Physical AI vs Robotics: Is This Really New?
Yes and no. Robotics industry is a well established industry, with a lot of fascinating research around it. You might remember that even OpenAI once had a robotics research arm. Back in 2019, its Rubik’s Cube-solving robotic hand captured headlines, demonstrating how reinforcement learning could teach machines to handle complex physical tasks. It felt like the beginning of something transformative – a seamless convergence of AI and robotics.
But in 2021, OpenAI quietly stepped away from robotics. The decision wasn’t due to failure but to a strategic pivot. Robotics demanded hardware integration, resource-heavy infrastructure, and solutions to the stubborn gap between simulated environments and the real world. For OpenAI, the returns didn’t justify the effort. They decided to double down on large language models like GPT – in 2021 GPT-3 was already showing a lot of potential for broader impact.
Thanks to that decision, we now have ChatGPT and the surge of Generative AI, which has elevated agentic systems to a completely new level (agents, as you might suspect, also carry a phenomenal history of development on their shoulders).
By enabling this leap in Agentic AI, OpenAI set the stage for others to reimagine what’s possible. Jensen Huang, inspired by this shift, envisioned the future of robotics enabled by advanced AI = Physical AI. Working backwards, he saw how NVIDIA could play a critical role in bringing that future to life, building the tools and infrastructure to make it happen.
Not a subscriber yet? Subscribe to receive our news digests and educational articles:
How Jensen Huang's "Working Backwards" Shapes NVIDIA's AI Strategy
Jensen Huang’s "working backwards" approach begins with envisioning the future of AI as a foundational layer across industries, where AI agents and physical systems act autonomously and intelligently. From this vision, he identifies the necessary elements – scalable architectures, robust training systems, and tools that Nvidia must develop to enable this future. Working backwards, Huang outlined Three Key Computers – the essential computing architectures pivotal for modern AI and robotics systems:
Training Computers designed to handle the massive data and computational power required for training AI models. These computers are essential for generating accurate and reliable AI models.
Simulation and Digital Twin Computers enable the creation of digital twins and simulations, providing a virtual environment to test and optimize AI and robotic systems before real-world deployment.
Deployment Computers are responsible for deploying trained models in real-world applications. Examples include AI computers in cars or robotics systems, which provide the computing power necessary for autonomous operations.
Why Robotics Is Becoming the Largest Computer Industry
The global workforce population is declining, and in some manufacturing-heavy countries, this decline is particularly significant. For these nations, ensuring that robotics becomes a productive and integral part of their industries has become a strategic imperative. With no foreseeable growth in their labor force, the urgency to develop and deploy advanced robotics systems has never been greater.
❝
Everything that moves will be autonomous. That’s a foregone conclusion. There's no limitations to robots. It could very well be the largest computer industry ever, and the reason for that is we don't need more cell phones than people. But, you know, robots, you can build as many as you like.
Jensen Huang
Industry-Specific AI Agents: From healthcare to logistics, specialized agents can revolutionize workflows and productivity.
Robotics and Automation: Physical AI has the potential to transform industries like manufacturing, warehousing, and autonomous transportation.
General-Purpose Robots: With advancements in synthetic data and imitation learning, the dream of versatile humanoid robots operating in unstructured environments is closer than ever.
NVIDIA's Physical AI Strategy: Infrastructure Over Products
One of the defining aspects of NVIDIA’s strategy, as articulated by Jensen Huang, is its commitment to shaping and enabling entire industries – not just creating individual products. NVIDIA operates like a tank with a built-in 3D printer: relentless in its progress, constantly constructing new realities in computation and AI. With every advance, NVIDIA reshapes the landscape, crafting frameworks that others rely on to innovate. Its vast resources drive simultaneous advances in hardware, software, and ecosystems, empowering industries to evolve and scale around its orbit. At the core of its strategy lies an ecosystem that equips developers, researchers, and enterprises with the tools to build on NVIDIA’s foundation, remaining the de facto choice for innovation.
They might or might not build their own robots, NVIDIA positions itself as a catalyst for scaling Physical AI across industries.
NVIDIA’s phenomenal partnership strategy
How Physical AI Economics Work: Deployment, Scale, Cost
The economics of human robotics, as described by Jensen Huang, revolve around ease of deployment, scalability, and leveraging economies of scale:
Brownfield Deployment Advantage: Human robotics stands out because it can be deployed in existing environments (Brownfield) without requiring significant changes to infrastructure. Unlike robots with wheels or tracks that often necessitate customized facilities (Greenfield deployment), human robotics can integrate seamlessly into current operations. This parallels the adoption of smartphones, which required no new infrastructure because they fit into users' existing lifestyles (e.g., pockets).
Ease of Deployment and Scalability: The simplicity of deploying human robots allows for rapid scaling. Once they are introduced into a facility or operation, they can be quickly replicated across various contexts. This scalability leads to increased adoption, driving down costs per unit as volume increases.
Economies of Scale and a AI Flywheel: As the deployment of human robots scales, economies of scale come into play. The cost efficiencies achieved by mass production and deployment fuel a technology flywheel, accelerating innovation. This cycle – where adoption drives cost reduction, leading to further adoption – ensures continual technological improvement and expanded applications.
Enabling Industries with Tailored Solutions: Nvidia focuses on empowering industries rather than competing directly. By providing tools like AI enterprise solutions, agentic AI toolkits, and modular hardware/software platforms, they enable diverse industries – automotive, robotics, and beyond – to integrate AI into their operations. This adaptability allows companies like Toyota, Waymo, and Tesla to tailor Nvidia’s stack to their specific needs, enhancing their own competitive edge.
Accelerated Innovation: The combination of easy deployment, scaling, and the flywheel effect means human robotics will see rapid advancements.
Nvidia predicts that within 10 years, these robots will reach a level of capability that surprises even skeptics. And as Jensen Huang sees it, it’s a multi-million industry.
NVIDIA's Long-Term Bet on Physical AI
NVIDIA’s investments in foundational technologies, such as Blackwell GPUs and AI simulation platforms, are designed to support industries for decades to come. This long-term perspective ensures that NVIDIA remains indispensable as AI adoption grows exponentially. By aligning its strategy with broader technological trends, NVIDIA ensures it’s not just part of the AI revolution but a leader in defining its direction.
And guess what? In January 2025, OpenAI announced its return to robotics. Caitlin Kalinowski shared job listings for hardware engineers and robotics specialists with experience in areas like autonomous cars, drones, wheeled robots, humanoids, and soft robots. The team aims to tackle the challenges that once led OpenAI to step back from robotics, now leveraging years of AI advancements and deeper industry insights. This shift highlights the growing potential for embodied AI.
Conclusion
So, where does that leave us? This year, AI agents are set to take off – starting with software engineering, digital marketing, and customer service. But Jensen Huang envisions a future where agentic and physical AI converge to redefine industries, from manufacturing to healthcare. For many countries with declining populations, relying on robots becomes imperative to maintain productivity and economic growth. From autonomous vehicles to industrial robotics, AI is quickly becoming the backbone of future industries.
These millions of new robots and autonomous vehicles will drive billions of dollars into supporting data centers. AI is no longer just a tool; we’re moving toward agents and assistants that use tools – a whole new layer above the computing stack. Every factory will have a digital twin, mirroring operations to optimize KPIs before real-world deployment. Huang’s warning is clear: if your enterprise isn’t leveraging AI assistants yet, you’re already falling behind.
OpenAI’s renewed focus on robotics only reinforces this shift. Their return signals that embodied intelligence is becoming a priority for the entire field, marking a critical step forward in the evolution of AI.
FAQ
What is the difference between agentic AI and physical AI?
Agentic AI operates primarily in digital environments, where AI agents can reason, plan, use tools, and complete complex workflows. Physical AI extends these capabilities into robots and autonomous machines that understand spatial relationships, physical dynamics, and real-world environments. In Jensen Huang's vision, agentic AI is the foundation that enables physical AI.
What are the four key characteristics of agentic AI?
According to Jensen Huang's framework, agentic AI combines four core capabilities: reasoning and planning, multimodal understanding, tool and data integration, and autonomous task execution. Together, these allow AI agents to solve complex problems with minimal human supervision.
Why did Jensen Huang call robotics the largest computer industry ever?
Huang argues that robots could eventually outnumber traditional computing devices because they can be deployed across manufacturing, logistics, healthcare, transportation, agriculture, construction, and homes. As AI becomes embodied, every autonomous machine effectively becomes a computer, creating a market potentially larger than smartphones or personal computers.
What are the real-world applications of agentic AI according to Huang?
Huang highlights AI research assistants, industrial optimization, software engineering, cybersecurity, healthcare, drug discovery, manufacturing, logistics, robotics, autonomous vehicles, and digital workers that automate complex business workflows. These applications serve as the bridge toward physical AI and intelligent robotics.
Bonus for our Premium members
What to Keep an Eye On as an Investor (10 Trends in AI and Robotics)
That’s all for today. Thank you for reading! Please send this newsletter to your colleagues if it can help them enhance their understanding of AI and stay ahead of the curve.