NVIDIA just shipped a safety system for robots. The map changed overnight.

On 22 June 2026, NVIDIA announced Halos for Robotics — described as the industry's first full-stack safety system for physical AI. It pairs safety-capable compute (the IGX Thor module, which carries an independent on-chip safety processor) with a safety software stack and an accredited inspection lab. More than forty partners signed on at launch, and Agility's humanoid Digit is the first robot to adopt it.

For two years, the claim that AI-driven robots need a dedicated hardware safety layer was a minority position — the kind of thing you had to argue for from first principles. The largest company in AI just moved it to the centre of the industry. Hardware safety for physical AI is no longer a fringe research topic. It is a product category.

I'll be direct: I argued this was inevitable. In May I published The Software Safety Ceiling — the case that physical AI cannot be made safe in software alone, because the substrate a software monitor runs on is the same substrate it is supposed to be watching. Aviation, nuclear and rail each reached that conclusion decades ago and mandated hardware-isolated safety. Robotics was next. NVIDIA just agreed, in the loudest way the industry has.

What NVIDIA got right is the architecture. Halos puts an independent safety processor — its own cores, its own power, its own clock — alongside the AI compute, rather than asking the AI stack to police itself. That is precisely the hardware safety layer the software-only camp said wasn't necessary. It turns out it is.

What the announcement leaves open is the harder question. 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. Is on-chip isolation the same thing as independence — or is it shared fate dressed up as separation? That question now decides how much any safety layer is actually worth, and it is the subject of the next post.

The category is set. The next race is over who provides the independent, vendor-neutral version of it.