Nvidia's Open Cosmos Bet Reframes Build-Versus-Buy for Physical AI -- Pure AI

Nvidia's Open Cosmos Bet Reframes Build-Versus-Buy for Physical AI

By John K. Waters
07/16/2026

Key Takeaways

Nvidia is distributing its Cosmos 3 world models, including the newly announced Cosmos 3 Edge, as open weights with training scripts and datasets, rather than as a closed service, positioning the platform layer itself as shared infrastructure.
A wide roster of Japanese manufacturers, from established robotics suppliers such as Fanuc and Yaskawa Electric to industrial groups such as Fujitsu and Hitachi, are adopting or contributing to the Cosmos framework rather than building proprietary world models from scratch.
The approach mirrors a broader pattern in enterprise AI, where vendors compete on infrastructure, tooling, and hardware integration around a shared open foundation rather than on model ownership itself.

Nvidia's introduction of Cosmos 3 Edge, and the roster of Japanese manufacturers adopting it, illustrates a build-versus-buy pattern taking shape in physical AI: rather than compete on proprietary world models, Nvidia is distributing the foundation model layer openly and competing instead on the hardware and tooling built around it.

Cosmos 3, the model family Edge extends, was released in May, with open weights, training scripts, and datasets published on Hugging Face and GitHub, according to Nvidia's launch announcement. Developers can customize the models using Hugging Face Diffusers and Nvidia NIM microservices, and access is also available through third-party cloud infrastructure partners including Microsoft Azure, CoreWeave, and Nebius, according to the same release.

Choosing Infrastructure Over Model Ownership

The companies named in Nvidia's July 16 announcement illustrate the pattern directly. Fujitsu is exploring development of a collaborative control platform for physical AI built on Nvidia's stack rather than an independent one, while Fanuc, Yaskawa Electric, and Kawasaki Heavy Industries are integrating Nvidia technologies into their existing robotics systems, according to Nvidia's release. A further group of companies, including Honda R&D, Omron, Telexistence, and Enactic, are applying the open Cosmos framework to specific use cases such as elder care robotics and retail automation, rather than developing their own foundation models for those applications, per the same release.

That pattern extends the logic Nvidia laid out when it first released Cosmos 3, describing the goal as unifying capabilities that developers previously had to assemble from separate models—world generation, controlled scene generation, scene understanding, and policy generation—into a single open foundation, according to Nvidia's developer blog covering the release. The technical rationale is that physical AI systems need to perceive, predict, and act using shared context, a requirement earlier, narrower Cosmos releases split across separate models and inference pipelines.

Where the Competitive Layer Shifts

With the model layer distributed openly, the competitive and commercial emphasis shifts to hardware and deployment tooling. Cosmos 3 Edge is built specifically for Nvidia's own edge computing hardware, including its RTX graphics processors, DGX systems and two newly announced Jetson modules, the T2000 and T3000. Nvidia also introduced new Metropolis software libraries alongside Edge, intended to help developers use coding agents to build and operate vision AI applications on top of the open model, per the same release.

That structure lets Nvidia capture value through hardware sales, cloud partnerships, and developer tooling, even as the underlying model weights remain freely available. It also lowers the barrier for manufacturers like Fanuc and Yaskawa Electric to adopt physical AI capabilities without independently training their own foundation models, a resource-intensive undertaking that Nvidia's own technical report describes as requiring large-scale multimodal training data spanning images, video, audio, and action sequences.

Caveats

Nvidia describes Cosmos 3 as an open platform, but that characterization warrants closer examination. Questions remain about the specific commercial licensing terms governing the Cosmos 3 model weights, because the term "open source" can encompass a wide range of usage rights. It is also not yet clear how deeply some members of the Cosmos Coalition, Nvidia's growing ecosystem of robotics companies, manufacturers, researchers, and software partners working to advance and adopt the Cosmos world model framework, contribute technically versus serving primarily as early adopters or ecosystem partners.

It is also worth noting that Nvidia's open-model strategy complements, rather than replaces, its core hardware business. Graphics processors, networking, and AI infrastructure remain the company's primary sources of revenue. Nvidia CEO Jensen Huang's comments that Japan helped build modern manufacturing and now has an opportunity to reinvent it were made in the broader context of encouraging adoption of Nvidia's computing platforms as the foundation for that transformation.

About the Author

John K. Waters is the editor in chief of a number of Converge360.com sites, with a focus on high-end development, AI and future tech. He's been writing about cutting-edge technologies and culture of Silicon Valley for more than two decades, and he's written more than a dozen books. He also co-scripted the documentary film Silicon Valley: A 100 Year Renaissance, which aired on PBS. He can be reached at [email protected].