Nvidia Boosts 'Physical AI' with New Blackwell Hardware & AI Models

At SIGGRAPH 2025, Nvidia unveiled a sweeping vision for what it calls “Physical AI,” a strategic convergence of artificial intelligence and computer graphics aimed at enabling systems to interact intelligently within the real world. This ambitious initiative encompasses everything from advanced robotics and autonomous vehicles to smart infrastructure, built upon a foundation of new hardware, sophisticated simulation platforms, and cutting-edge AI models.

“AI is advancing our simulation capabilities, and our simulation capabilities are advancing AI systems,” explained Sanja Fidler, Nvidia’s Vice President of AI Research, underscoring the symbiotic relationship at the heart of this strategy. The company’s comprehensive ecosystem is designed to deliver the computational power and intelligence necessary for these real-world applications.

Central to Nvidia’s push are new hardware offerings based on the Blackwell architecture, tailored for demanding AI workloads. For data centers, the Nvidia RTX PRO 6000 Blackwell Server Edition GPU is set to integrate into mainstream enterprise servers, leveraging the widely adopted 2U form factor. Major system partners, including Cisco, Dell Technologies, HPE, Lenovo, and Supermicro, are slated to offer these servers. Nvidia states these systems represent a significant leap from traditional CPU-based architectures to accelerated computing platforms, boasting up to 45 times higher performance and 18 times better energy efficiency compared to CPU-only configurations. The new GPUs feature fifth-generation Tensor Cores that support the FP4 format, a key innovation Nvidia claims increases inference performance sixfold over the previous-generation L40S GPU.

For the desktop segment, Nvidia introduced two compact graphics cards: the Nvidia RTX PRO 4000 SFF Edition and the RTX PRO 2000 Blackwell. These cards are engineered to bring AI acceleration to smaller, more energy-efficient workstations, catering to professionals in fields like engineering, design, and 3D visualization. The RTX PRO 4000 SFF reportedly delivers up to 2.5 times higher AI performance at the same 70-watt power consumption as its predecessor, while the RTX PRO 2000 is said to offer 1.4 times faster performance in Computer-Aided Design (CAD), among other improvements. Both new GPUs are expected to become available later this year.

This powerful new hardware provides the computational backbone for Nvidia’s Physical AI vision, particularly its emphasis on simulation. The core idea is to create highly realistic, physically accurate digital twins where AI systems, such as robots, can safely learn through extensive trial and error before deployment in the physical world. “Computer graphics and AI are converging to fundamentally transform robotics,” affirmed Rev Lebaredian, Vice President of Omniverse and Simulation Technologies at Nvidia.

The technological bedrock for this simulation-first approach is the Nvidia Omniverse and Isaac platforms. Nvidia announced new software libraries for Omniverse, including Omniverse NuRec, which facilitates the reconstruction of real-world environments from sensor data using advanced 3D Gaussian splatting techniques. Additionally, the robotics simulation applications Isaac Sim 5.0 and Isaac Lab 2.2 are now available as open-source projects on GitHub, incorporating these new rendering capabilities.

A compelling real-world application of this simulation-first strategy comes from Amazon Devices & Services for its “zero-touch” manufacturing process. Here, CAD models of new products are imported into Nvidia Isaac Sim to generate over 50,000 synthetic images. These images are then used to train AI models that control robotic arms, enabling them to autonomously perform quality checks or integrate new products into the production line. This entire process relies purely on skills learned in simulation, eliminating the need for physical hardware modifications. Technologies like the FoundationPose pose-estimation model further empower these robots to recognize even novel objects without prior training.

To ensure AI systems can not only perceive but also reason effectively, Nvidia has expanded its AI model families. For enterprise applications, the Nemotron family now includes Nemotron Nano 2 and Llama Nemotron Super 1.5. These models are designed to enable AI agents to tackle complex, multi-step tasks across sectors like customer service and cybersecurity. Nvidia highlights the models’ high efficiency, achieved through a hybrid architecture and quantization (NVFP4). Companies such as CrowdStrike, Uber, and Zoom are reportedly already testing or planning to integrate these models.

Specifically developed for Physical AI is Cosmos Reason, a customizable 7-billion-parameter Vision Language Model (VLM). This model is engineered to empower robots and vision AI agents to interpret and act within the physical world by incorporating prior knowledge, an understanding of physics, and “common sense.” Its applications span robot planning, automated annotation of training data, and video analytics. Uber, for instance, is utilizing Cosmos Reason to analyze the behavior of autonomous vehicles, while VAST Data and Milestone Systems are employing it for intelligent traffic monitoring.

To translate these advanced technologies into tangible applications for intelligent infrastructure, Nvidia integrates many of these components into its Metropolis platform. The platform has been enhanced with several new features, including the seamless integration of the Cosmos Reason VLM, new vision foundation models within the TAO Toolkit, and extensions for Isaac Sim to generate rare training scenarios. Partners are already leveraging Metropolis for diverse solutions. Accenture and Belden are developing “smart virtual fences,” simulated in Omniverse, to enhance worker safety around industrial robots. DeepHow is using the Metropolis VSS blueprint for a “Smart Know-How Companion” that transforms work instructions into visual guides, a solution that Anheuser-Busch InBev reportedly used to reduce onboarding time for new employees by 80 percent.

Nvidia’s “Physical AI” initiative represents a comprehensive effort to bridge the gap between digital simulation and real-world intelligence, promising a future where machines learn and act with unprecedented autonomy and understanding.