One booth at Embedded World in Nuremberg stayed in my mind long after the show ended.
Not because it was loud.
The opposite.
No giant AI slogans. No AGI claims. No "future of humanity" presentation.
Just a dark wall with five words:
"FPGA & SoC. Made in Germany."
The company was Trenz Electronic — a relatively small German embedded systems company, founded in 1992, AMD Alliance Partner since 2007, over 25 years building FPGA modules, SoMs, and custom electronics. Not a household name. Not a unicorn. Not on anyone's disruption watchlist.
And most visitors at Embedded World probably walked past without stopping.
But standing there, I had a specific thought:
This booth is showing exactly what the global AI conversation keeps misunderstanding about Europe.
What was on the wall
The Trenz product display was organized as a wall of modules — each on its own pedestal with a touch-screen kiosk for detailed specifications. Reading across the labels:
→ Zynq UltraScale+ SoMs — multiple variants, AMD's heterogeneous SoC combining ARM Cortex-A53 application processors with UltraScale+ FPGA fabric. Real-time processing, hardware acceleration, industrial-grade reliability → Versal AI Edge SoM — AMD's newest architecture, integrating AI Engines (vector processors) directly alongside programmable logic and ARM cores. Not "AI" as a marketing word — AI as a silicon architecture decision → Spartan-10 SoM and Carrier — the cost-optimized FPGA family for high-volume industrial applications where power and BOM matter more than raw performance → TE0803 MPSoC Series — multi-processor SoC modules for applications requiring concurrent real-time and application-layer processing
Every one of these products represents a design decision made years before it appears on a show floor. The Versal AI Edge architecture alone required AMD to redesign its silicon from the ground up to integrate AI inference engines alongside traditional programmable logic.
Trenz is not building these chips. They are building the modules that make these chips usable in real industrial systems — handling PCB layout, signal integrity, thermal management, power sequencing, and long-term supply chain reliability. The engineering that disappears into the product.
The second booth: Edge AI Vision at Scale
Twelve meters away, a different kind of demo was running.
A large display wall — approximately 3×2 meters — showing twelve simultaneous camera feeds. Live. Real-time. Each feed a fisheye or wide-angle view, capturing the show floor, faces, objects, motion.
The banner above: "Edge AI Vision at Scale."
Beside it: "Cloud Development."
On the bench: a compact embedded board — not a rack server, not a GPU cluster. A board approximately the size of a paperback book, with a camera module attached.
To the right of the bench: a small RC car, wired with sensors and a camera, running autonomously on a test track.
The right-side screen showed a terminal — active code, log output, real-time inference results scrolling.
This was the same show, the same hall, the same hour as booths showing 100,000-GPU clusters and trillion-parameter model benchmarks.
And here: twelve camera streams, real-time inference, autonomous vehicle demo — on hardware you could carry in a backpack.
What this reveals about the AI conversation
If you follow global technology headlines, the conclusion feels obvious:
→ America dominates AI foundation models → China scales infrastructure aggressively → Europe is falling behind
At the narrative level, this is largely accurate.
Europe does not control the leading LLMs. It does not own hyperscale cloud infrastructure. It does not move at Silicon Valley speed. It does not have the venture capital density to compete with San Francisco.
These are real structural disadvantages.
But the show floor at Embedded World looked more complicated.
Because the AI revolution people discuss online — and the engineering systems that actually run the physical world — are not always the same thing.
Physical systems obey different rules
A large language model running in a data center can tolerate: → Variable latency (a few hundred milliseconds is acceptable) → Temperature-controlled environment → Redundant power and cooling → Software update cycles measured in weeks → Failure modes that produce wrong text
A factory robot cannot tolerate any of these.
It needs: → Deterministic timing — the actuator must respond within microseconds, every time → Operation at -20°C to +70°C ambient → EMC immunity from motors, inverters, and switching power supplies → Functional safety certification (IEC 61508, ISO 26262 for automotive) → Hardware availability guarantees of 10–15 years → Failure modes that are defined, predictable, and fail-safe
These are not software problems. They are not problems you solve by scaling compute.
They are physical-system engineering problems — and they require a completely different set of skills, tools, and institutional knowledge than training foundation models.
Europe's actual position
Europe is genuinely weak in: → AI foundation model development → GPU infrastructure ownership → Hyperscale cloud platforms → Consumer internet platforms → Venture capital speed and scale
Europe remains genuinely strong in: → Industrial automation and control → Power electronics and drives → Embedded real-time systems → Functional safety and certification → Machine vision for industrial applications → FPGA-based hardware acceleration → Edge AI deployment in physical environments → Precision measurement and metrology
The European industrial embedded systems market is still projected to grow from approximately €6 billion in 2025 to nearly €8 billion by 2030 — driven by IoT integration, automation, and edge AI deployment. Not because Europe is winning the AI narrative. Because physical systems keep expanding.
The Versal AI Edge question
The most technically interesting product on the Trenz wall was the Versal AI Edge SoM.
AMD's Versal architecture is a fundamental departure from traditional FPGA design. Instead of adding "AI support" as a peripheral — the way most FPGA vendors initially approached neural network acceleration — Versal integrates AI Engines as first-class silicon elements alongside the programmable logic fabric and ARM cores.
The AI Engines are VLIW (Very Long Instruction Word) vector processors, each capable of 128-bit SIMD operations at high clock rates. Multiple AI Engines work in parallel, connected by a high-bandwidth network-on-chip. The result is an architecture that can run neural network inference at tens to hundreds of TOPS — not in a data center, but in a module the size of a credit card.
This is what "Edge AI" means when it is done properly:
Not a cloud service with an edge label. Not a microcontroller running a tiny quantized model. But a purpose-built silicon architecture for running real inference workloads at industrial power and temperature specifications.
Whether Versal will become the dominant Edge AI platform — or whether it will be displaced by purpose-built inference chips from startups and Asian manufacturers — is an open question.
But it represents a specific engineering philosophy: hardware designed to run AI reliably in the physical world, not to benchmark impressive numbers in a lab.
What "Made in Germany" still means
The Trenz booth headline was not a nationalist statement. It was a supply chain and reliability claim.
For industrial customers with 10-year product lifetimes, "made in Germany" means: → Known manufacturing provenance → Export control compliance → Long-term support commitment → Engineering team accessible in the same timezone → Certifications applicable to EU markets
These are not exciting engineering problems. They are the engineering problems that make industrial products possible to deploy and maintain.
And they require deep institutional knowledge that cannot be replicated quickly by scaling a software team.
The misunderstanding this reveals
The internet increasingly treats AI as if technological supremacy belongs entirely to whoever trains the largest models.
But technological strength has always had multiple dimensions.
The countries and companies that dominated the industrial revolution were not the ones with the best theoretical physics. They were the ones who could reliably manufacture steam engines, precision instruments, and chemical processes at scale.
The current AI revolution will transform physical systems — factories, vehicles, energy infrastructure, medical devices, agricultural equipment. That transformation requires both:
→ The AI capabilities being developed in American and Chinese research labs → The physical-system engineering capabilities concentrated in European and Japanese industrial companies
Neither is sufficient alone.
The Trenz booth was not showing the future of AI.
It was showing the infrastructure layer the future of AI will have to run on.
That is easy to underestimate when you spend most of your time reading about model benchmarks.
Much harder to underestimate when you spend time on a factory floor.
The booth that didn't look futuristic
Standing in front of that wall of FPGA modules, the feeling was not excitement about the next paradigm shift.
It was something more specific:
Recognition that the engineers who built these modules — and the tens of thousands of engineers who will design them into industrial systems over the next decade — are solving problems that foundation models cannot solve, cloud services cannot replace, and venture capital cannot shortcut.
Europe is not winning the AI narrative.
But the physical world still has to be engineered.
And that work is still being done.
Quietly. Stubbornly.
In Nuremberg, and in factories across Europe.
All photos: Thomas · @SignalByThomas · Embedded World, Nuremberg
