The short version of the misunderstanding: Omniverse gets treated as a rendering tool — the thing that makes the demo videos look good — and not much else. That’s not what it is, and the gap between the two is bigger than a rendering engine.
What’s actually underneath the image
Omniverse is not one application. It’s a stack, and the renderer is one layer out of six: hardware at the bottom, then the core data model and simulation engines (OpenUSD, PhysX, Newton, Warp), then the platform layer (Kit, the extension framework), then cloud APIs, then the actual applications people click into (Isaac Sim, USD Composer, DRIVE Sim), and finally reference blueprints that wire several of those together for a specific job — a robot fleet, a data center twin, a factory line. The RTX renderer sits in the engine layer, third from the bottom. It’s necessary. It is not the platform.
The clearest way I’ve found to make this concrete: NVIDIA’s framing for Physical AI describes three different computers you need, not one. A DGX system trains the robot’s brain. An RTX server simulates the world the robot will operate in — builds the digital twin, generates synthetic training data, runs the physics. A small Jetson computer riding on the actual robot runs the trained policy in real time, on hardware, at sensor frame rates.
Omniverse is the middle one. Not the picture at the end of the pipeline — the simulation environment that makes the other two computers’ jobs possible at all. You can’t train a robot by trial and error on real hardware; it’s too slow, too dangerous, and too expensive to fail physically thousands of times. So you fail in simulation, millions of times, cheaply, and only bring the policy that survived simulation onto real hardware. Omniverse is where that failing-cheaply happens. The render is what that place looks like in a screenshot.
Digital twins work the same way, minus the robot. A twin isn’t a 3D model you rotate on a screen — it’s a live simulation you run “what if” scenarios against before committing to something physical and expensive: reconfiguring a factory floor, testing a wireless network’s coverage before the towers go up, routing a robot fleet through a warehouse that doesn’t exist yet. The visual is the artifact you’d show a stakeholder in a meeting. The simulation underneath it is the actual product.
This is my own individual opinion, formed from my personal experience working with this technology. It does not represent an official NVIDIA position. I may have gotten something wrong or missed a detail — always happy to be corrected.