Nvidia used SIGGRAPH 2026 to hand AI agents the keys to its 3D design software, folding new Omniverse libraries into its Agent Toolkit so autonomous systems can build and inspect simulation-ready worlds on their own.
Nvidia doesn't just want to sell you the chip anymore. It wants to sell you the agent that runs on it. That's the substance behind Nvidia's SIGGRAPH 2026 agentic AI tools announcement this week in Los Angeles, and it shows exactly how far up the software stack the company is now willing to climb.
The centerpiece is an expansion of Nvidia's Agent Toolkit, the open framework the company uses to give AI agents access to its own technology. As of this week, that toolkit includes new Omniverse libraries, ovrtx for RTX sensor simulation, ovphysx for GPU-accelerated physics, and a set of CAD-to-SimReady skills, all published openly on GitHub, according to a release covered by StockTitan. The goal is narrow but real. An AI agent can now open a 3D scene, run a physics test, and convert a raw CAD file into something ready for simulation, without a person driving the software by hand.
Six companies are already building on it. SideFX, PTC, ForgeCAD, Lightwheel, Moonlake AI and Palatial are among the first adopters, Nvidia said. Palatial is using the CAD-to-SimReady skills to automate asset validation straight from CAD inputs. Lightwheel's SimReadyGen tool, built on OpenUSD, generates physically accurate simulation-ready assets from a text prompt instead of a modeler's hours of manual work.
That's the whole pitch in one line. Type a sentence, get a usable 3D asset.
The rest of the industry piles in #
The access isn't confined to Nvidia's own software, either. SideFX is bringing Model Context Protocol support to Houdini 22 through a new APEX Script workflow, letting AI assistants read APEX syntax, documentation and examples to help artists write procedural character rigs. Epic Games has done something similar, opening Unreal Editor to AI clients through MCP so agents can interact with editor functions directly, GamesBeat reported from the event. MCP is becoming the common tongue.
Nvidia also introduced Cosmos Reason, a 7 billion parameter vision-language model built for physical AI and robotics. It's meant to let a robot or a vision system reason the way a person does, drawing on physics understanding and plain common sense to decide what to do next in the real world. That's the pitch, anyway. Nvidia says its broader Cosmos world foundation models have already been downloaded more than 2 million times.
The real fight is over the software layer #
None of this is happening in a vacuum. Nvidia has spent 2026 watching the ground shift beneath the layer it built its empire on. Cheap, open-weight Chinese models, Kimi K3 and Qwen3.8 among them, have spent recent months undercutting the assumption that frontier AI capability requires frontier-priced compute. If a startup can get most of the way there on an open model that costs a fraction of what a closed system runs, the chip alone stops being the moat.
That's the real logic behind this week's push. Selling GPUs is still Nvidia's core business, and it isn't going anywhere. But selling the software layer agents actually run on, Omniverse, Cosmos, NIM microservices, locks developers into Nvidia's ecosystem in a way a commodity chip never could. AMD is chasing the same hardware market with its own accelerators. Google has been shipping its own TPU silicon alongside Gemini. Neither has anything close to Omniverse's install base in professional 3D and simulation tooling, built up over more than a decade of work with studios, engineering firms and now robotics startups.
Frankly, that's the harder moat to break. A rival chip can match Nvidia on raw throughput eventually, the way AMD's MI series already competes on paper. Rebuilding a decade of studio relationships and asset pipelines around a competitor's agent framework is a different problem entirely, and it's the one Nvidia is betting the open-weight wave can't solve for its rivals.
Nvidia hasn't published pricing for running these agents at scale in production. The libraries themselves are free and open on GitHub today. What isn't free is the compute an agent burns running physics simulations and generating assets around the clock, and that compute still runs, for now, on Nvidia hardware.
Also read: Moonshot's Bet on a Bigger Kimi K3 Is Paying Off in China's AI Race • Jeff Bezos and the UK Government Just Bet $450 Million on CuspAI • Google Shares Rise After Report of a Chip Designed Around Gemini