NVIDIA AI agents for local models and simulation

NVIDIA has published examples of developers directing frontier AI models to assemble simulation applications. The work covers assets, physics, rendering and checks that a scene behaves as intended. Developers give instructions in natural language, then review results and steer changes. Omniverse libraries supply GPU-accelerated physics, rendering and sensor simulation. The agent does the assembly; the developer keeps the judgement. That matters if you build tools rather than just use them: the output stays editable in existing artist software, so you can inspect and refine it.

Nick Scarcella, a senior technical artist on NVIDIA's BondLab team, guided agents to recreate an OP-1 portable synthesizer from product images and manuals. He exported the model into OpenUSD, tested it in Houdini and Blender, then refined details before adjusting materials for rendering. A separate project by Pradeep Kaushik built a what-if application around a simulated gas turbine. Users can change load, ambient temperature and fuel composition, then compare power, efficiency and fuel consumption. Six test procedures cover scenarios including a stuck fuel valve and an exhaust-sensor error. These are prototypes for exploring scenarios, not proof of real-world performance.

NVIDIA announced the SimReady Central Library at GTC Berlin. It offers more than 2,000 simulation-ready USD assets from equipment manufacturers and asset providers, aimed at robotics simulations and digital twins. Paul Kind, from NVIDIA's Omniverse documentation team, pointed an agent at the library and assembled a kitchen scene through conversation. The scene held 180 placements drawn from 63 unique assets, plus 18 custom geometry pieces. He reviewed renders from Isaac Sim and guided adjustments. The catch is scope: the assets carry embedded simulation properties, so you still need to pick ones that fit your task and test them in an assembled environment.

NVIDIA argues that larger models, longer contexts and agentic applications are raising GPU memory pressure while supply tightens. Its answer is not simply more memory per chip. The company designs GPUs, CPUs, networking, storage and software together to get more work out of each byte. HBM is the fastest and most contested tier, with model weights and KV cache competing for space. Below it sit CPU memory, local SSDs and shared storage, each trading speed for capacity. Bandwidth, not capacity alone, sets the ceiling. Once memory bandwidth cannot feed compute fast enough, extra HBM adds cost and power without helping.

NVIDIA reports that HBM capacity per GPU rose roughly 3% from Blackwell Ultra to the Vera Rubin platform. Over the same step, agentic throughput per dollar rose by up to 30x on the SemiAnalysis AgentX workload. On a DeepSeek V4 Pro workload, GB300 NVL72 is claimed to deliver up to 10x higher throughput per dollar than the H200 platform at 100 tokens per second per user. At 125 tokens per second per user, Vera Rubin NVL72 is claimed at up to 30x over GB300 NVL72. These are vendor figures tied to a named benchmark, so treat them as claims to check against your own workload, not as a promise for your setup.

Chris Alexiuk, a Toronto-based senior product research engineer in deep learning at NVIDIA, argues that transparency is a precondition for good AI, not a feature. He says users should have access to the weights, the training data and the training recipes. He points developers to the Nemotron family of open multimodal models, built for long-running and specialised agents. His split is practical: a smaller model such as Nemotron 3.5 Lightning for repetitive, high-frequency work, and Nemotron Ultra for long-running autonomous agents and deep research. The limit is that this is one engineer's view, not a measured comparison against closed models.

If you run models locally, the useful question is how much context you can keep resident and how many users you can serve at a target speed. NVIDIA's answer leans on shrinking weights and KV cache to 4-bit precision, spreading weights across up to 72 GPUs, and moving idle context to cheaper storage. Several model builders have released NVFP4 checkpoints, including Black Forest Labs, Thinking Machines Lab and Meta. None of this tells you what your own hardware will do. I have not tested these setups, and the throughput figures come from NVIDIA and SemiAnalysis. Start from your latency target, then check whether the memory tiering fits your machine.