At the SIGGRAPH 2026 conference in Los Angeles, running through July 23, NVIDIA made a series of major announcements spanning agentic AI, physical AI, synthetic media detection, and desktop supercomputing.
The most significant shift is the opening of creative tools to AI agents through the Model Context Protocol (MCP). Adobe, Canva's Affinity, Blender, Boris FX, Foundry, SideFX Houdini, and Epic Games' Unreal Engine are all adopting MCP connections, allowing AI agents to inspect scenes, flag missing textures, generate playblasts, build node trees, and validate shots against pipeline rules — all while keeping creative control in human hands. NVIDIA's RTX PRO workstations, DGX Spark, and DGX Station systems bring local AI performance to artists and studios, reducing reliance on cloud services.
On the physical AI front, NVIDIA released Cosmos 3 Edge, a 4-billion-parameter omnimodel now available on Hugging Face under an open license. Optimized for NVIDIA Jetson, RTX PRO, and GeForce RTX GPUs, Cosmos 3 Edge ranks No. 1 on VANTAGE-Bench in its parameter class and enables real-time vision analytics, robot action, and autonomous vehicle scene understanding on device. It is part of the broader Cosmos 3 family (Edge 4B, Nano 16B, Super 64B), all now openly available.
NVIDIA also announced the Synthetic Video Detector NIM microservice, which analyzes video frame by frame to produce a classifier score detecting synthetic content. The model achieves up to 92% accuracy on uncompressed video and can process 1080p in as little as 22 milliseconds on RTX systems. Wowza is embedding the microservice into its Video Intelligence Framework for real-time detection in livestreaming workflows.
For AI agents on the desktop, NVIDIA unveiled the full stack for DGX Station: the NVIDIA Agent Toolkit combining NemoClaw open blueprints, Nemotron 3 Ultra (a 550B-parameter open model), Omniverse libraries, and the OpenShell secure runtime — all running on the GB300 Grace Blackwell Ultra Desktop Superchip with up to 20 petaflops of FP4 AI compute. Setup takes roughly 30 minutes, and multiple DGX Stations can be linked for distributed workloads. LangChain and Nous Research have already tuned their agent harnesses for Nemotron 3 Ultra.
NVIDIA also presented 21 accepted technical papers at SIGGRAPH covering neural rendering, physics simulation, and world models. Notably, MotionBricks — a real-time motion model trained on over 350,000 motion clips — can simultaneously drive animated characters and a Unitree G1 humanoid robot, demonstrating the convergence of graphics and physical AI.
