Physical AI Without Location Is Like LLMs Without the Internet

Last week, Jensen Huang stood on stage at GTC and said something worth sitting with: every industrial company will become a robotics company.

He's right. But left unmentioned in this declaration is a critical tech problem that nobody on that stage has yet solved - except for ZaiNar.

You can seek to build the best robot foundation model in the world. You can train it in photorealistic simulation, on petabytes of synthetic data, across millions of virtual environments. And the moment you deploy it in an actual industrial environment (a mine, a port, an active warehouse) it doesn't usefully know where it is without significant added hardware, software and compute. Think about what made large language models possible. It wasn't just the models. It was the internet - the most comprehensive, continuously updated record of human thought ever assembled: millions of articles, billions of social media posts, near-infinite content.

LLMs didn't become powerful just because the algorithms got clever. They became powerful due to clever algorithms trained on massive amounts of quality data from real use cases.

Physical AI needs the same thing. Not simulated movement. Observed movement: how a pushcart gets navigated through busy aisles; how humans move through a complex, risky industrial site; how forklifts handle blind corners; or how cross-functional hospital staff, patients and visitors move around an entire ICU. What’s needed is continuous, real-time, sub-meter spatial data, not just models - where every object is, how fast it's moving, how the environment around it is changing. That data layer doesn't exist yet at any meaningful scale.

That's what ZaiNar builds.

We turn existing 5G networks into spatial sensing systems - no new hardware, no cameras, no satellites. The signals are already there, transmitting billions of times per second. We synchronize them at sub-nanosecond precision and derive location at sub-10cm accuracy, indoors and outdoors, at ranges up to 1.5 kilometers, updating 100 to 500 times per second.

Critically, because our system operates entirely on the network side, it works without any software on the device being located - no battery impact, no compute overhead, no dependency on Apple or Google granting access.

We announced commercial availability of this system three weeks ago. Since then, we have been in daily meetings with carriers, industrial operators, and autonomous systems companies who are asking the same question in different languages: where is everything, right now, all the time?

The robot foundation models are ready. The simulation frameworks are ready. The compute is ready.

Physical AI is waiting for the data layer. We built it - and we have been building it for nine years, for exactly this moment.

Notes

Unlike conventional positioning approaches that rely on dedicated positioning reference signals (PRS), ZaiNar uses the Sounding Reference Signal (SRS) that every device already transmits - 100 to 500 times per second versus PRS's and PSRS’s maximum of once per second. This is why ZaiNar can track fast-moving objects, and why it requires no cooperation from device operating systems.