Can Nvidia's physical AI strategy replicate its AI dominance?

Nvidia's three-computer architecture

At the center of Nvidia's physical AI strategy is what the company calls its three-computer architecture, which involves training, simulating and deploying physical AI systems.

The first part of this architecture is the computing infrastructure used to train AI models. The second part uses Nvidia's Omniverse and Cosmos platforms to simulate physical environments, generate data and test AI systems before deploying them in the real world. The third part brings computing to the edge, where platforms such as Jetson run AI workloads inside robots and other autonomous systems.

"It is foundational because Cosmos and Omniverse address two of physical AI's biggest bottlenecks: data and validation," Docca said. "Cosmos helps generate and reason over diverse scenarios, while Omniverse lets developers test them safely and at scale in physically based simulation." These technologies, combined with Nvidia's accelerated computing for model training and real-time edge deployment, create a platform spanning the physical AI lifecycle, which Docca described as a key competitive advantage for Nvidia.

The simulation component of Nvidia's three-computer architecture is already being adopted in industrial settings. Foxconn, HD Hyundai, PepsiCo and Kion Group are using Siemens' Digital Twin Composer, which incorporates Nvidia Omniverse libraries, to model manufacturing and logistics operations before deployment.

The edge computing piece of the architecture moves AI out of the data center and into the machines themselves. Nvidia expanded that part of its portfolio in August with Jetson Orin Nano 2, an entry-level robotics computer that runs AI workloads at the edge. Nvidia is targeting the platform for robots, delivery and inspection drones and vision AI systems. The module and developer kit are expected to be available in the first half of 2027.

Nvidia's data center strategy pairs freely available software with software that runs most effectively on Nvidia hardware, Ray noted. He sees the company attempting to replicate that model at the edge with physical AI. "The three-computer architecture provides Nvidia with a reason to expand beyond the data center into the edge where there is more competition," he said.

Solving physical AI's data problem

Training physical AI systems is different than software-based AI training. Robots and autonomous machines need data that reflects how objects, people and environments behave in the physical world, including situations that can be difficult, expensive or dangerous to reproduce.

Nvidia unveiled its Physical AI Data Factory Blueprint in March, an open reference architecture that automates training data generation, augmentation and evaluation. The goal is to reduce the cost, time and complexity of producing large amounts of data to train physical AI systems.

The blueprint brings together several Nvidia technologies. Cosmos Curator processes and organizes real-world data, Cosmos Transfer generates variations and synthetic scenarios and Cosmos Evaluator assesses generated data. Omniverse provides physically based simulation. The system also uses Nvidia's Cosmos world foundation models and coding agents to turn limited amounts of training data into larger synthetic data sets. Included is data representing rare events and other scenarios that can be difficult or impractical to capture in the real world.

The blueprint is another way Nvidia is expanding its role beyond the computing hardware used to train and run physical AI systems. Hexagon Robotics, Teradyne Robotics and Uber are among the physical AI developers using the blueprint, according to Nvidia. The company is also building tools to help developers create and prepare the data that physical AI systems need.

Building Nvidia's physical AI ecosystem

Nvidia's physical AI strategy depends on expanding its technology stack and building an ecosystem of companies that use and integrate its technologies. Rather than building most of the finished machines, Nvidia is providing technology that robotics companies and other partners can use to develop and deploy their own systems.

"The platform will evolve as the market advances, but our core strategy remains to enable a broad ecosystem to build the next generation of physical AI applications and systems," Docca said. That ecosystem is particularly important in physical AI because robots and other autonomous systems require technologies beyond AI computing alone.

[O]ur core strategy remains to enable a broad ecosystem to build the next generation of physical AI applications and systems.

As part of the Halos safety ecosystem, Lattice Semiconductor is developing Halos-certified designs using Nvidia's Holoscan Sensor Bridge, which connects high-bandwidth sensor data to AI computing systems. "Lattice is proud to be investing significantly into [Nvidia's] ecosystem," said Raemin Wang, Lattice's vice president of segment marketing. "At the same time, the scope of the opportunity is far broader than any single company."

Physical AI systems combine AI models and computing with sensors, networking, safety systems, robotics platforms and edge infrastructure; But these components require specialized technologies, Wang explained, adding that "no single provider can deliver every component or every area of expertise."

As part of an expanded collaboration announced in August, Amazon Robotics is using Nvidia's Jetson platform, Omniverse libraries and Isaac robotics development platform to develop next generation robots. The work spans simulation, synthetic data generation, robot training, route optimization, functional safety and real-to-sim validation.

Nvidia is also moving further into humanoid robotics. In May, the company introduced its Isaac GR00T Reference Humanoid Robot. The open reference design combines a Unitree humanoid robot with Nvidia's Jetson Thor computing and Isaac GR00T software and models. It illustrates Nvidia's broader approach to the robotics market. Rather than manufacturing the entire robot, Nvidia is providing computing, models and development software that can combine with hardware from other companies. Research institutions, including Stanford University, ETH Zurich and the University of California San Diego, will use the reference design for humanoid robotics research.

Nvidia's physical AI challenges

Nvidia's dominance in AI data centers does not guarantee the company will achieve the same position in physical AI. Robots and other autonomous machines require a broader mix of computing technologies as AI moves into real-world environments. "[A]s you move out of the data center and onto a factory floor or into a robot, the architecture has to extend beyond what any single processor can cover alone," Wang said.

GPUs remain the backbone of physical AI, handling the training and inference workloads that make these systems possible, Wang said. But physical AI systems must also process continuous streams of data from cameras, LiDAR and other sensors and coordinate that information with inference and real-time control.

"Most real-world failures in physical AI don't happen inside the model," Wang explained. "They happen at the boundaries between sensing, inference and control. Those are system-level hardware problems, and solving them requires chips that enforce contracts on timing and data format, not just accelerate computation. The real question isn't whether one processor is better than another. It's about placing the right workload on the right processor."

Wang said he sees physical AI developing around heterogeneous architectures, with GPUs handling computationally intensive AI workloads, while specialized processors handle functions such as sensor aggregation, connectivity and deterministic real-time control. ​"Qualcomm is already in this market, with Hailo [Technologies] and Intel having market presence," Gartner's Ray said, "so, establishing a de facto standard based around vertically integrated hardware will be challenging. Having said that, Jetson is well placed, so I wouldn't say it's impossible, just more difficult."

Nvidia will rely on third parties to provide this orchestration, but those third parties may want to see their orchestration platforms being portable across edge hardware.

Integration with existing enterprise systems presents another challenge. It'll be "critical, as physical AI must work with enterprise resource management platforms and even building control," Ray said, noting that a mobile robot might need to coordinate with elevators, doors and fire alarms. "Nvidia will rely on third parties to provide this orchestration," he said, "but those third parties may want to see their orchestration platforms being portable across edge hardware."

Yet another challenge is moving physical AI systems from testing into production. "The gap between a successful pilot and a production deployment is where physical AI systems get tested," Wang said. "Labs are controlled. Production environments are not."

Unlike software-based AI, physical AI failures can have severe real-world consequences. Safety, latency and system reliability are therefore critical considerations as robots and autonomous systems move into workplaces. "A robot that hesitates at the wrong moment because of an unexpected latency spike isn't just a performance problem, it's a liability," Wang said.

Cybersecurity is another concern. Physical AI systems haven't been deployed long and widely enough to be extensively tested by malicious actors, Wang said, warning that companies should not interpret the relative lack of known attacks as evidence that the systems are secure.

Above all, however, the cost and disruption of deployment will be "the most important factor," Ray said. Companies will need to determine where robots fit into existing workflows and what changes are required to safely accommodate them, he added. That could include separating robots from parts of a factory floor or providing additional training for employees.

Nvidia acknowledges some of these physical AI challenges. Physical AI systems need diverse and physically accurate data, as well as better transfer from simulation to real-world environments, Docca said. Companies also need systems that can integrate with existing technology, meet safety and cybersecurity requirements and demonstrate measurable returns. "Adoption will scale application by application as those conditions are proven," Docca reasoned.

Is a physical AI ecosystem possible?

For enterprise technology leaders, Nvidia's expanding physical AI ecosystem creates opportunities to build on a growing set of integrated technologies. But that raises questions about how dependent organizations should become on any single technology provider.

"Technology leaders should avoid designing physical AI architectures around a single processor, interface or deployment model," Wang advised. "The technology landscape is evolving rapidly, and the most resilient systems will prioritize flexibility, interoperability, security and adaptability from the beginning."

Technology leaders should avoid designing physical AI architectures around a single processor, interface or deployment model.

Flexibility could become increasingly important as physical AI deployments mature. Unlike software that can be replaced relatively quickly, robots and other physical AI systems could remain in service for years as standards, cybersecurity requirements and AI capabilities continue to change. Architectures that can evolve without major redesigns will be better positioned to manage cost and risk, Wang said.

The same applies to the software that connects physical AI with enterprise and industrial systems. Third-party providers responsible for that orchestration might want their platforms to remain portable across different edge hardware rather than tied to a single vendor, Ray noted.

Nvidia enters this market with significant advantages, including its position in AI computing, its developer ecosystem and a physical AI platform that now stretches from training and simulation to edge computing, robotics software and safety. But physical AI is also a more heterogeneous market, where Nvidia will need to coexist with other processors, hardware providers, robotics companies and industrial technology platforms.

Whether Nvidia can turn that broad strategy into the kind of dominance it achieved in AI data centers remains an open question. For enterprises, the more immediate challenge will be determining where physical AI can deliver measurable value while maintaining the flexibility, safety and interoperability needed for systems that operate in the real world.

Liz Hughes is an award-winning editor and writer covering AI and emerging technology and the former editor of AI Business and IoT World Today.