When Physical AI Meets the Network

While analyzing the many developments in the field of AI, one thing quickly becomes clear: the conversation is no longer just about AI. More and more it’s about the infrastructure that allows intelligence to move from the digital world into the physical one.

Some weeks ago, Rony Cohen from Flolive and I had the opportunity to speak in Amsterdam. We took the opportunity to share our views on how this new infrastructure changes the traditional IoT paradigm.

The discussion started with a familiar assumption. When we think about AI, we tend to focus on models, algorithms, and computing power. But once AI begins to control robots, machines and autonomous systems, another layer becomes critical: the network. Its role goes far beyond connectivity, influencing how data moves, where it is processed, and how reliably intelligence translates into action.

IoT has historically been about connecting sensors, collecting data, and sending this data to a central platform. Physical AI introduces a much more demanding scenario. Machines must continuously sense their environment, interpret data, and take action. Increasingly, those decisions need to happen in real time. All this puts the network under immediate pressure.

The network becomes part of the intelligence

As our conversation developed, the relationship between AI and the network became increasingly apparent: a sophisticated AI model is of no value if the data it requires arrives too late. Latency, jitter, packet loss, and network availability can influence how effectively a physical system responds to its environment. The discussion therefore moves beyond familiar telco topics like bandwidth, availability and coverage. The real architectural question nowadays is where intelligence should reside.

Some processing may need to happen directly on the device. Other workloads can be handled at the edge, while more computationally intensive tasks can remain in the cloud. The resulting architecture is distributed across device, network, edge and cloud.

This is where edge computing becomes particularly relevant. By processing data closer to the source, organizations can reduce latency, minimize network traffic, and potentially increase resilience when cloud connectivity is interrupted.

Intelligence creates new responsibilities

The conversation then touched on another important dimension: security and data sovereignty.

When machines become intelligent and connected, organizations need greater control over their data flows. Where is data generated? Where is it processed? What priority does AI-generated data have compared to other data? Which information can leave the local environment? How is the connection secured? These are no longer questions for the IT department. They become functional, architectural, and strategic decisions.

A different way of designing Physical AI

What stands out from our exchange is the need to reverse the traditional approach.

Rather than selecting connectivity after the AI application has been designed, connectivity should be considered part of the AI architecture from the beginning. Therefore, the important question is not simply which network technology to deploy. It’s what the Physical AI application actually requires from its network: latency, reliability, resilience, security, scalability, and data sovereignty.

For me the discussion ultimately lead to a simple but powerful conclusion: Physical AI may be powered by intelligent algorithms, but it can only operate as effectively as the infrastructure connecting intelligence to action. The smarter machines become, the more intelligent—and intentional—our networks will need to be.