Physical AI Needs an Industrial Nervous System

At GTC 2026, NVIDIA’s chief executive told the industry that “every industrial company will become a robotics company.”¹ Robot foundation models, synthetic-data simulation, and the first humanoid pilots on real production lines, including humanoid assembly cells in German plants,² have moved Physical AI from the keynote stage to the factory floor. The race to own this shift is on.

Yet almost all of the attention, and almost all of the capital, points at three layers: the silicon, the models, and the robots. The layer that actually decides whether Physical AI works inside a running factory gets very little of either. Physical AI does not become real through better GPUs or larger models, but an industrial nervous system: the runtime that connects perception to action, deterministically, in real-time, on the machines you already own.

1. From algorithm to organism

If a robot model is a brain, a GPU is the metabolic energy that powers it. Neither is of any use to a living organism without a nervous system: the tissue that carries signals from the senses to the muscles fast enough, and reliably enough, to act in the physical world. Industrial Physical AI behaves the same way. The intelligence is necessary, but it is inert until something carries its decisions into deterministic, real-time motion on real hardware.

The mapping from biology to the factory is direct:

A foundation model can recognize a defective part or plan a trajectory. That value is realized only if the perception data arrives with deterministic timing, the control decision executes inside the machine cycle (microseconds, not the 200 milliseconds typical of a cloud dashboard), and the result is delivered over the real fieldbus to a real actuator. Latency that is invisible in a chatbot is catastrophic on a filling line or a robot arm. Physical AI is therefore a real-time, deterministic integration problem first, and a model problem second.

2. The missing layer

The market has organized itself cleanly around three of the four layers of Physical AI. NVIDIA owns the hardware: silicon, edge compute, simulation. The foundation-model labs own the intelligence: perception, planning, language. The robotics companies own domain-specific training and reasoning: the embodied skills and task logic that make a particular robot good at a particular job, proven in impressive but controlled demonstrations. The fourth layer, industrial runtime infrastructure, the software that makes AI run natively and deterministically inside the control loop of a real machine, is almost unowned.

This is not a small gap. NVIDIA itself has started reaching toward it, with edge orchestration and industrial-grade edge compute,³ which is a signal of how strategic the runtime is. But orchestrating AI workloads at the edge is not the same as controlling a machine in real time. Today’s industrial edge and IoT platforms can move and visualize data, yet they cannot close a control loop at two milliseconds. Today’s PLC platforms can close the loop, but they cannot host an AI model in the same runtime that drives the machine. The classical PLC, the workhorse of the last forty years, was never built to think, but to continuously repeat the same instructions.

That is the structural opening. The industrial AI market was already worth USD 43.6 billion in 2024 and is growing at roughly 23 percent a year,⁴ and whoever provides the runtime that is open, real-time, hardware-agnostic, and AI-native, owns the layer on which all the others depend.

3. The brownfield reality: add AI without replacing everything

There is a practical reason this layer matters more than the demos suggest. Real factories cannot be rebuilt. A typical plant runs PLCs, SCADA systems, and machines that are expected to last ten to fifteen years and represent enormous sunk investment. Rip-and-replace is not an option for almost anyone. The realistic path to Physical AI in brownfield environments is additive: layer modern software, edge AI, data pipelines, and orchestration on top of the controllers and machines that are already in place.

In practice that means keeping the certified safety and motion subsystems exactly where they are, and adding a real-time software layer alongside them that can read existing fieldbuses, normalize the data semantically, run AI inference locally, and feed decisions back into the process. The proof already exists: in one brownfield packaging deployment, the approach cut customization effort by roughly 60 percent without replacing any core infrastructure. The road to Physical AI does runs through modernizing existing factories, not starting new on a greenfield.

4. The twin nervous system: simulation and the real world

A real nervous system does more than react. It maintains an internal model of the body and its surroundings, and uses that model to predict and rehearse a movement before a single muscle contracts. Industrial Physical AI needs the same faculty. That internal model is the digital twin: a simulated nervous system running in parallel with the physical one, exercised safely in software before, and alongside, the real machine.

The bridge between the two is FMU/FMI, the open standard that lets simulated and physical components speak the same language. With it, control logic can be co-simulated against virtual machines and validated long before it touches real hardware; mechanical design and control software can be developed in parallel rather than in sequence; and the twin can stay synchronized with the running line to predict outcomes, optimize behavior, and de-risk every change. The payoff is concrete: virtual commissioning compresses time-to-value, and in test-heavy environments such as aerospace test beds it can cut manual test effort by roughly half.

5. Where Xentara fits

Xentara, built by embedded ocean GmbH, is the runtime for this missing layer: an open, real-time, hardware-agnostic software platform that unifies control, connectivity, semantics, and AI inference in a single runtime. Its defining capability is simple to state and rare to find: AI running natively inside the real-time control cycle, on any standard hardware, without a cloud dependency. Cameras and other sensors feed AI models, and those models drive control output in the same cycle that runs the machine, at cycle times measured in microseconds on standard COTS hardware under Linux RT.

That is precisely the nervous system the metaphor calls for:

• Perception: native vision and sensor ingestion with semantic data abstraction
• Reflexes: deterministic real-time control at the edge, with machine learning inside the loop
• Coordination: a modular, skill-based architecture of 50+ skills orchestrating control, connectivity, and AI as containerized microservices
• Signal pathways: native southbound connectivity (EtherCAT master, OPC UA, PROFINET, Modbus, S7) and northbound links to REST, databases, and cloud
• Foresight: a digital-twin-ready architecture with FMU/FMI co-simulation, so control logic is validated against a virtual machine before it runs on the real one

Because it is hardware-agnostic and brownfield-ready, Xentara layers onto existing PLCs and SCADA rather than replacing them, and it deliberately leaves certified safety to specialized SIL-rated subsystems. Early results from machine builders are concrete, not theoretical:⁵

The runtime is the moat

Conventional wisdom says the value in Physical AI will accrue to whoever owns the smartest model or the fastest chip. The history of computing suggests otherwise: value concentrates in the layer that everything else has to run through. In industrial Physical AI, that layer is the runtime, the nervous system that turns perception into deterministic action on real machines, green- and brownfield. NVIDIA owns the hardware. The model labs own the intelligence. The robotics companies own domain-specific training and reasoning. The industrial nervous system is still open, and it is the layer that decides whether any of the rest ever reaches the factory floor. That is the layer Xentara is built to provide.

About Xentara

Xentara is developed by embedded ocean GmbH, based at Steinerstrasse 15 in Munich. Positioned as The Physical AI Enabler, it is a software-defined automation platform that brings deterministic real-time control, open IT and OT connectivity, semantic data abstraction, and native AI inference into a single runtime, on any standard hardware. It is built for both greenfield and brownfield, layering onto the machines you already operate.
Learn more about Xentara, or reach out to our team to talk through what an industrial nervous system would look like on the machines you're already running. 

References

1. NVIDIA GTC 2026; Jensen Huang, “every industrial company will become a robotics company.” Manufacturing Dive, 2026. ↩
2. NVIDIA, AI-driven manufacturing at Hannover Messe 2026; Hexagon and the AEON humanoid performing assembly at a BMW plant in Leipzig, one of the first humanoid deployments in a German production line. ↩
3. NVIDIA launched its OSMO edge orchestration framework in January 2026. Source: ZEDEDA, 2026 industrial edge predictions. ↩
4. Global industrial AI market USD 43.6 billion in 2024, projected ~23% CAGR through 2030. Source: ZEDEDA, 2026. ↩
5. Source: Xentara customer and project references (Syntegon, Liebherr Aerospace, BenThor automation GmbH) and an anonymized proof-of-concept with a European specialty-machinery OEM. xentara.io ↩