Three superpowers, one missing layer — what mapping Physical AI revealed

For the last month I did one thing: I mapped the places that actually build Physical AI. Not the places that talk about it — the places where the robots, the talent, and the money are. There are three that matter at scale, and they could not be more different from one another. Looking at them side by side told me something the regulation and the standards never did.

Israel is velocity. Around 123 Physical AI companies across perception, autonomy, defense, and agri-robotics — an ecosystem that turns research into deployed hardware faster than anywhere else, with unusual density in the sensing and autonomy layers that sit closest to the AI. What it is thin on is the same thing it is thin on everywhere: a safety-certification layer. Fewer than a fifth of the companies I mapped have a visible certification strategy. The velocity is real. So is the gap underneath it.

Europe is depth and law. Eight clusters, each built on decades of engineering heritage: German industrial robotics and humanoids, the Danish collaborative robot that essentially created the category, Swiss precision, French defense-and-aerospace crossover, Italian automation, the Spanish market that is now the continent's fourth-largest, British embodied AI and surgical robotics, and Dutch logistics. Europe is also where the rules live — the EU Machinery Regulation that applies from January 2027 — and where the certification houses live: TÜV, Pilz, UL. Everything you would need to assemble an independent safety layer is on this continent. It just hasn't been assembled into one.

Japan is philosophy. Japanese manufacturers supply roughly 38% of the world's industrial robots, the single largest national share, and five of the ten largest robot makers on the planet are Japanese. But the number that stayed with me wasn't a market-share figure. At the iREX exhibition in December 2025, every AI-enabled robot on the floor — FANUC's, Yaskawa's, all of them — shipped with hard-coded safety overrides the AI itself cannot bypass. Not as a premium option. As a baseline. In Japan, hardware-enforced safety is treated as an inherent property of a well-made machine, not an external compliance burden. No regulation mandates it. The culture does.

Three ecosystems. Three strengths — velocity, depth, philosophy. Three completely different theories of how to build a robot. And the same hole runs straight through all three: none of them has an independent safety layer. Each builds intelligence brilliantly. None had built the thing that sits between the AI and the existing safety infrastructure — the certification houses, the safety silicon, the safety PLCs that all already exist — and does it for robots specifically, independently of whoever made the AI.

For two years that was a theoretical gap, the kind of thing you argue for from first principles. Then, on 22 June 2026, it stopped being theoretical. NVIDIA announced Halos for Robotics — described as the industry's first full-stack safety system for physical AI, pairing safety-capable compute with a safety software stack and an accredited inspection lab. More than forty partners signed on; Agility's humanoid Digit is the first to adopt it. This matters, and it deserves to be said plainly: the largest company in AI just moved hardware safety for robots from a fringe position to the center of the industry. The category is now real.

But read the architecture carefully. NVIDIA's safety processor lives on the same chip, comes from the same vendor, and is designed by the same team as the AI it supervises. It runs on NVIDIA's own compute. That is a genuine step forward — and it is also exactly the question functional safety has answered the same way for fifty years. Aviation, nuclear, and rail all concluded that the highest integrity levels require independence and diversity from the system being judged, not just isolation inside it. A safety monitor that shares a vendor, a chip, and a design team with the thing it monitors is isolation. It is not yet independence.

So the verdict from three maps, on three continents, is not the one I expected to write when I started. It is not "the safety layer is missing." Not anymore. It is sharper than that: the safety layer now exists — and it belongs to one company. The independent, vendor-neutral version — the one that works across robot platforms instead of welding to a single vendor's stack, the one that can serve as a genuinely diverse second channel rather than a part of the system it is checking — is still nobody's product. Israel has the velocity to build it. Europe has the regulation and the certification houses to demand it. Japan has the philosophy that already assumes it. None of them has shipped it.

That is what the map actually shows. Not an absence anymore — a contest. The most important safety layer in Physical AI has exactly one occupant, and the lane next to it, the independent one, is wide open.

I turned this into a book. Physical AI Safety: The Missing Layer — Why Physical AI needs an independent safety layer, and how to certify it is the long-form version of everything above: the safety math worked by hand, real incidents taken apart, the EU Machinery Regulation read clause by clause, and a certification framework rendered as a working spec. It's available now on Amazon — Kindle today, paperback in a few days.