NVIDIA CTO Michael Kagan on AI Factories, Physical AI, Sovereign Intelligence and the Future of Computing in an Exclusive Interview with Dinis Guarda

NVIDIA Chief Technology Officer Michael Kagan joins Dinis Guarda in an exclusive interview at the NVIDIA GTC Berlin 2026 to discuss the major computing revolutions, the transformation of data centres into AI factories, the economics of tokens per watt, the rise of physical AI and robotics, and why countries and enterprises cannot afford to outsource their intelligence.

In an exclusive episode of the Dinis Guarda YouTube Podcast, Michael Kagan, Chief Technology Officer of NVIDIA, joins Dinis Guarda, author, entrepreneur, and founder of Businessabc, to explore the technological foundations of the AI revolution and its implications for businesses, governments, industries, and society.

Drawing on more than four decades of experience in processor architecture, networking, and high-performance computing, Kagan shares insights from his career at Intel, Mellanox, and NVIDIA. He discusses how computing has evolved from individual processors to interconnected data centres, why artificial intelligence is creating unprecedented demand for computational resources, and how physical AI could open an entirely new chapter in technological development.

“Computing requirements started to grow every quarter with the pace of 10x a year instead of 2x every year. Computing evolved from a single chip to the hunger scale machine - that's what requirements came and that's what we need to deliver.”

This architectural evolution has major implications for the future of AI infrastructure.

Modern AI systems require thousands of processors to operate together, exchanging information across complex networking environments.

Michael highlights the importance of predictable communication between these processors, particularly when large numbers of machines must coordinate their operations.

“One of the key characteristics of old parallel computing is predictability of execution time jitter. And the network jitters are killing you if you do not design the network carefully because what happens is that you have multiple machines, thousands, sometimes hundreds of thousand machines that are doing the job and then they start to exchange the results to go to the next step.”

AI Factories and Tokens per Watt: A New Economic Model for Intelligence

As artificial intelligence becomes embedded in business operations, NVIDIA is increasingly describing AI infrastructure through the concept of AI factories.

In traditional industrial production, factories transform raw materials into finished goods. In AI factories, computing systems process data and generate tokens that support intelligent applications.

“Tokens are the basic element of intelligence and you need to out of these tokens you basically build your solution. And for the AI factories, power is the main thing. The main constraint of building the factory is the power. So the more intelligence you can produce per unit of power, the better off you are and the better the factory you're building.”

Physical AI: The Next Frontier Beyond Digital Intelligence

Michael argues that the computational requirements of physical AI could eventually exceed those associated with today's digital AI applications.

Discussing the scale of the opportunity, he states:

"Whatever we have now is nothing compared to what's going to physical. You need a place to for robots which is the physical AI to gain the experience. Once you have this experience in the robot, now you have a computer that trains the model and then you have a computer to train the robot to simulate the environment to simulate the real world.”

Simulation-based training could enable autonomous systems to develop experience without exposing people, equipment, or infrastructure to unnecessary risks during the learning process.

However, Michael also explains that robots must be capable of making decisions locally.

“Everything that moves needs to be autonomous and it cannot rely on being able to always connect to the data center. So it needs to be able to make local decisions completely autonomously.”

OpenShell and the Challenge of Governing Autonomous AI Agents

As AI systems evolve from generating information to executing tasks independently, governance becomes an increasingly important technological requirement.

During the interview, Dinis Guarda asks Michael about NVIDIA OpenShell and its approach to policy enforcement for autonomous AI agents.

Michael compares the governance of AI systems with the rules and safeguards established within human societies.

“What we are doing at NVIDIA is we developed a platform that enables our customers' ecosystem to define the rules they need and this platform will help them to enforce these rules. We don't believe we can succeed without others. Once it's open then the ecosystem can innovate and enable everyone to customize this platform to fit these needs.”

AI as an Amplifier of Human Ambition

Beyond the technological architecture of AI, Michael discusses its potential to expand human capabilities.

"AI is one of the biggest ... amplifiers of human capabilities probably since the invention of the wheel. I have never seen anybody in the engineering world that has enough resources. When we get more resources, we have more desires and more ambitions."

This perspective challenges the assumption that AI's primary economic value lies in replacing existing work.

Instead, AI could enable individuals and organisations to undertake activities that were previously impractical because of limitations in time, expertise, or resources.

For business leaders, this raises an important strategic question: should AI investment focus primarily on reducing costs, or on expanding the range of products, services, and opportunities that organisations can create?