オムニバースへ：オープンワールドモデルがPhysical AIの最前線を切り開くInto the Omniverse: How Open World Models Push the Frontier of Physical AI

Physical AI, the branch of artificial intelligence that enables machines to perceive, reason about and act within the physical world, is emerging as one of the industry's most closely watched frontiers, and NVIDIA is arguing that openness is the way to advance it. In July, the company joined more than 200 companies and organizations in signing "Open Weights and American AI Leadership," an open letter contending that AI leadership will ultimately be measured not by any single frontier model but by whether an open ecosystem reaches every sector.

The letter reframes how progress in AI should be judged. Rather than concentrating capability inside a handful of proprietary, closed systems, the signatories argue that broad and distributed access to open-weight models will determine which industries and economies actually benefit. "Open weights" refers to models whose trained parameters are published, allowing developers to inspect, fine-tune and deploy them on their own infrastructure rather than depending solely on an interface controlled by a single vendor. That distinction matters for regulated or safety-critical fields, where organizations often need to run models locally, audit their behavior and adapt them to specialized tasks.

For NVIDIA, this position connects directly to its work on Physical AI and its Omniverse platform, a system for building and operating physically accurate 3D simulations. The company's broader thesis is that so-called open-world models, which learn the dynamics of environments and can generate realistic scenarios, are a prerequisite for teaching robots, autonomous vehicles and other embodied systems how the physical world behaves. Because gathering real-world training data at scale is slow, expensive and sometimes dangerous, simulation and synthetic data have become central to the field. A model that can predict how objects move, collide and respond to forces can be used to train and test machines long before they are deployed on a factory floor or a public road.

World models, sometimes described as open-world models in this context, aim to capture that understanding. NVIDIA has previously introduced world foundation models under its Cosmos initiative, designed to generate simulated video and environments for training Physical AI systems, and it maintains related tooling through its Isaac robotics stack and Omniverse. The company's argument appears to be that releasing such models with open weights would let a wider range of developers build on shared foundations rather than rebuilding capabilities from scratch, which is likely to accelerate adoption across manufacturing, logistics, healthcare and transportation.

The open letter also carries a policy dimension, framing open models as tied to national competitiveness under the banner of "American AI Leadership." That framing reflects an ongoing debate over whether open or closed approaches better serve both innovation and security. Proponents of open weights point to transparency, reproducibility and the ability for smaller organizations to participate, while critics raise concerns about misuse and the difficulty of controlling models once their weights are public. The signatories fall on the side that a vibrant open ecosystem, reaching every sector rather than a single leading lab, is the more durable measure of leadership.

This move fits within a wider industry shift. Open-weight model families from multiple developers have proliferated over the past two years, and hardware vendors have an incentive to encourage them because open models expand the total base of developers building and running workloads, much of it on accelerated computing. NVIDIA's business is closely tied to that demand, so its support for open weights aligns with its commercial interests as well as its stated technical vision, a context worth keeping in mind when weighing the argument.

For readers tracking robotics and simulation, the practical takeaway is that the boundary between digital and physical AI is narrowing. Tools that generate synthetic environments, digital twins that mirror real facilities, and models that understand physical dynamics are increasingly treated as parts of a single pipeline from training to deployment. Whether an open ecosystem ultimately delivers the reach the letter envisions will depend on how models are licensed, how accessible the supporting tools remain, and how quickly industries beyond early adopters put them to use. The signing itself signals where NVIDIA and a large group of peers believe the field is heading.