NVIDIA Omniverse is a GPU-powered platform for building, simulating, and collaborating inside realistic, physics-based 3D worlds — used for robotics, digital twins, and AI-driven system design.
NVIDIA Omniverse is built directly on top of CUDA and related NVIDIA SDKs. It’s essentially a “metaverse operating system” powered by accelerated computing. It’s realized by foundational CUDA packages:
ey CUDA-based packages and SDKs behind Omniverse
| Layer | CUDA SDK / Toolkit | Role in Omniverse |
|---|---|---|
| Rendering | OptiX, RTXGI, RTXDI | GPU ray tracing, global illumination, and light sampling; built on CUDA kernels. |
| Physics Simulation | PhysX (CUDA-accelerated) | Simulates rigid body, fluid, and cloth physics in real time. |
| AI & Deep Learning | TensorRT, cuDNN, CUDA-X AI, PyTorch w/ CUDA | For AI-assisted denoising, upscaling, animation, and neural material generation. |
| Data Processing | cuDF, cuML, cuGraph, RAPIDS | GPU-accelerated dataframes, ML, and analytics used inside Omniverse extensions (especially for robotics and digital twins). |
| Networking / Multi-GPU | NCCL, CUDA-Aware MPI | For distributed GPU simulation or rendering (essential in large collaborative scenes). |
| 3D Frameworks | USD (Universal Scene Description) + CUDA-based pipelines | The “metaverse language” used by Omniverse, accelerated for GPU memory and scene streaming. |
| Simulation AI | Isaac Sim (built on Omniverse) | Uses CUDA and PhysX to simulate robots and environments in real time. |
In NVIDIA GTC 2025, Jenson Huang mentioned AI factories many times, “These aren’t data centres. These are factories that manufacture intelligence.” nextdc.com+2NVIDIA Blog+2
“AI is now infrastructure … just like the internet, just like electricity, needs factories.” NVIDIA Blog+1
In a transcript snippet:
“This is how intelligence is made, a new kind of factory generator of tokens, the building blocks of AI.”, AI factory” refers to a new way of thinking about infrastructure for artificial intelligence: large‑scale, industrialized compute facilities that produce intelligence rather than simply storing data or serving workloads.
Data center as a cost center is outdated , the AI factory is treated more like a production line which generates outputs (intelligent capabilities) that can be monetised. It’s not just hardware; the value lies in software, data pipelines, model libraries, domain expertise — making the “factory” concept broader than just a server farm.
Naixian Zhang 