Nvidia used the SIGGRAPH 2025 conference to broaden its push into "physical AI"—systems that perceive, reason and act in the real world—by unveiling new foundation models and Blackwell-based processors for data centres and compact workstations. The headline software release is Cosmos Reason, a 7-billion-parameter vision-language model designed to give robots and video-analytics agents the ability to interpret scenes and plan next steps. It joins updated Nemotron Nano 2 and Nemotron Super 1.5 models and plugs into Nvidia’s Metropolis, Omniverse and Isaac Sim toolkits to streamline synthetic-data generation, digital-twin simulation and edge deployment. On the hardware side, the RTX PRO 6000 Blackwell Server Edition—equipped with 96 GB of GDDR7 and delivering up to 3.7 petaFLOPS of FP4 compute—has been validated for 2U rack servers from Cisco, Dell, HPE, Lenovo and Supermicro. Nvidia says the configuration can run physical-AI and generative-AI workloads up to 45 times faster, with an 18-fold energy-efficiency gain, compared with CPU-only 2U systems. Shipments start 2 September. For engineers and creators who need local inference, Nvidia introduced two small-form-factor workstation cards: the RTX PRO 4000 SFF and RTX PRO 2000. Both fit within a 70-watt power envelope; the higher-end model promises up to 2.5-times the AI throughput of its predecessor while maintaining half-height, dual-slot dimensions. By coupling the Cosmos models with a full range of Blackwell GPUs—scalable from pocket-sized workstations to energy-efficient servers—Nvidia is positioning itself as a one-stop supplier of the compute, software and data required to bring autonomous factories, smart cities and other physical-AI applications to market.