On this page
The ecosystem and infrastructure
The previous chapter asked a structural question: will humanoid robotics be dominated by vertically integrated companies, or will it evolve into an open...
The previous chapter asked a structural question: will humanoid robotics be dominated by vertically integrated companies, or will it evolve into an open ecosystem where thousands of companies build on shared platforms? In January 2025, NVIDIA provided a decisive answer.
At CES (Consumer Electronics Show), Jensen Huang declared that "the ChatGPT moment for robotics is coming." He announced Cosmos, a platform of world foundation models designed to democratize physical AI. Within months, Cosmos had been downloaded over 3 million times. By September, NVIDIA had released Isaac GR00T N1.6, the latest in a family of open foundation models for humanoid robots that brought together reasoning and action in a single system.
This was more than a product launch. It was the opening salvo in what may become the defining structure of the humanoid robotics industry.
The Hourglass Architecture#
The most useful framework for understanding the emerging industry structure is the hourglass architecture, similar to the structure that enabled the internet's explosive growth.
At the top of the hourglass sits a vast diversity of applications. Hundreds of companies are building robots for specific use cases: warehouses, hospitals, homes, farms, construction sites, eldercare facilities. Each requires different capabilities, different form factors, different price points. No single company can serve all these markets effectively.
At the bottom sits a competitive landscape of hardware manufacturers and component suppliers. Some focus on advanced actuators. Others on lightweight materials. Still others on batteries, sensors, or mechanical designs optimized for specific tasks. Chinese manufacturers like Unitree are driving costs down dramatically, while precision engineering firms in Germany and Japan optimize for performance.
In the middle, at the narrow waist of the hourglass, sits what might become the Humanoid Robot Core: a standardized set of interfaces, shared foundation models, common operating systems, and agreed-upon communication protocols that allow the top and bottom to interoperate freely.
This architecture mirrors the internet's success. The internet thrived not because one company controlled everything, but because TCP/IP (Transmission Control Protocol/Internet Protocol) and HTTP (Hypertext Transfer Protocol) created a thin waist that let millions of applications run on millions of devices. The value was created at the edges, not at the center.
The Robot Operating System (ROS) has long provided basic services for heterogeneous robotic systems: hardware abstraction, device control, message-passing between processes. With ROS 2 now the recommended platform following ROS 1 Noetic's end of support in May 2025, the foundation exists for something more comprehensive.
The question is whether the humanoid robot core will extend beyond ROS to include world foundation models, shared training data, common APIs (application programming interfaces) for manipulation and navigation, and standardized form factors that make it economical to build components at scale. NVIDIA is betting it will.
The Open Ecosystem Takes Shape#
NVIDIA's strategy reveals why open ecosystems may ultimately prevail. Training world foundation models costs hundreds of millions of dollars and requires infrastructure that only a handful of companies possess. By making these models available under open licenses, NVIDIA gifts the robotics community capabilities that most companies could never develop alone.
Why would NVIDIA do this? Because their business model depends not on selling robots but on selling the infrastructure that everyone building robots will need. Cosmos models were trained using thousands of NVIDIA GPUs through DGX Cloud. Make robotics accessible to thousands of companies, and all of them become customers for NVIDIA hardware and compute services.
The results have been remarkable. At GTC (GPU Technology Conference) in March 2025, NVIDIA announced Isaac GR00T N1, the world's first open, fully customizable foundation model for generalized humanoid reasoning and skills. The model uses a dual-system architecture inspired by human cognition: a fast-thinking action model for reflexive responses and a slow-thinking reasoning model for deliberate planning. By May, NVIDIA had released GR00T N1.5 at COMPUTEX, improving the model's ability to adapt to new environments and recognize objects through user instructions. In September, GR00T N1.6 arrived at the Conference on Robot Learning in Seoul, integrating Cosmos Reason, an open reasoning vision language model that lets robots break down complex instructions using prior knowledge and common sense. Cosmos Reason topped the Physical Reasoning Leaderboard on Hugging Face, downloaded over 1 million times.
The synthetic data capabilities are equally transformative. Using the Isaac GR00T Blueprint, NVIDIA generated 780,000 synthetic trajectories, the equivalent of 6,500 hours or nine continuous months of human demonstration data, in just 11 hours. Combining this synthetic data with real-world data improved model performance by 40%. The open-source NVIDIA Physical AI Dataset on Hugging Face has been downloaded over 4.8 million times, providing thousands of synthetic and real-world trajectories for developers worldwide.
This cuts to the heart of why open ecosystems may win. Real-world data is the currency of AI progress. The company that gathers the most varied, high-quality data across the widest range of tasks and environments will train the best models. But here is the challenge for closed systems: no single company can deploy robots into every setting where humans work. A vertically integrated company must focus on specific high-value markets.
An open ecosystem faces no such limitation. Hundreds of specialized companies, each focused on one application domain, one customer segment, one geographical region, can all contribute data and improvements back to shared foundation models. The sum total may far exceed what any single company can collect.
The industry is responding. Companies including Agility Robotics, Boston Dynamics, Disney Research, Figure AI, Foxconn, Franka Robotics, Hexagon, LG Electronics, Lightwheel, Mentee Robotics, NEURA Robotics, Skild AI, Solomon, Techman Robot, and XPENG Robotics are adopting NVIDIA's Isaac and Omniverse technologies. NVIDIA's Jetson Thor, powered by a Blackwell GPU (graphics processing unit), has been adopted by Figure AI, Galbot, Google DeepMind, Mentee Robotics, Meta, Skild AI, and Unitree for real-time on-robot inference. Boston Dynamics' adoption is particularly significant: the company that set the benchmark for dynamic robotics over two decades has chosen to build on open platforms rather than go it alone.
At the Conference on Robot Learning in Seoul in September 2025, NVIDIA technologies were referenced in nearly half of the accepted papers, with adoption across leading research labs including Stanford, Carnegie Mellon, ETH Zurich, the University of Washington, and the National University of Singapore.
The Full-Stack Alternative#
Not everyone is embracing the open model. Tesla continues to pursue vertical integration, controlling everything from silicon to software to service. Chapter 8 examined Tesla's advantages in detail: AI infrastructure trained on millions of vehicles, battery expertise, manufacturing scale, and the ability to optimize across every layer of the system.
The economics of this model mirror Apple's success in smartphones. Build advanced technology in-house. Scale production to drive down costs. Capture premium margins not just from hardware sales but from ongoing software and services. Apple commands 15-20% of global smartphone unit sales but captures over 40% of industry profits.
Tesla is betting it can replicate this in robotics. In November 2025, the company began construction on a dedicated Optimus manufacturing facility at Gigafactory Texas, with a target of 10 million units annually by 2027. Drone footage showed ground clearing and site preparation for what would be the first purpose-built humanoid robot factory at scale. Meanwhile, a pilot line at the Fremont Factory aims to reach 1 million units annually by late 2026, with Optimus Gen 3 prototypes expected to debut in early 2026. Musk called Optimus "the biggest product of all time," with the potential to exceed automotive in long-term value. If Tesla executes, it could capture the premium segment while the open ecosystem serves everyone else.
But the full-stack model demands massive upfront investment, world-class expertise across multiple domains, and years of patient capital. Few companies can pursue this path. Those that succeed may capture outsized returns, but the barriers to entry are formidable.
The Physics Revolution#
One development deserves special attention. In March 2025, NVIDIA announced a collaboration with Google DeepMind and Disney Research to develop Newton, an open-source physics engine purpose-built for robotics. By September, Newton had entered beta and was released to all developers, managed by the Linux Foundation.
Newton addresses a fundamental challenge: robots must understand physics to interact safely with the real world. They need to know that cups can spill, that doors have hinges, that surfaces have friction. Previous physics engines were designed for games or scientific simulation, not for training robots that will operate alongside humans.
Built on NVIDIA's Warp framework and OpenUSD, Newton is optimized for robot learning and compatible with simulation frameworks like Google DeepMind's MuJoCo and NVIDIA Isaac Lab. With Newton's flexible design and ability to work with different physics solvers, developers can now simulate complex robot actions, like walking through snow or gravel and handling cups and fruits, and successfully deploy them in the real world.
Research labs at ETH Zurich Robotic Systems Lab, Technical University of Munich, and Peking University are already adopting Newton. Robot companies like Lightwheel and Franka Robotics are building on it. The physics engine is becoming part of the shared infrastructure that defines the narrow waist of the hourglass.
Security and Sovereignty#
Not every industry will accept dependence on a single platform provider, however open its licenses may be. Security and defense sectors demand auditable code, verifiable supply chains, and the ability to customize deeply for national requirements.
This creates opportunity within the open ecosystem model. Governments can build on open foundation models and contribute improvements back to the global commons while maintaining control over specific implementations deployed in sensitive contexts. The expensive foundation, physics models, training data, core algorithms, can be shared. Control over critical deployments remains sovereign.
The lesson of GPS (Global Positioning System) has been learned. No major power wants to repeat that dependency with humanoid robots, which may become even more fundamental to economic and military capabilities than satellite navigation. China's Ministry of Industry and Information Technology (MIIT) called for a "full-stack humanoid ecosystem by 2025" precisely because strategic autonomy requires domestic capabilities at every layer.
The open model paradoxically enables sovereignty by making the foundational layers accessible to all nations. Build on shared infrastructure. Differentiate on implementation. Maintain control where it matters most.
What This Means#
Market projections for humanoid robotics vary wildly, from $6 billion to over $50 billion by 2035, depending on methodology and assumptions. Goldman Sachs projects $38 billion by 2035 with 1.4 million units shipped. Yole Group projects $51 billion. Morgan Stanley extends further, projecting $5 trillion by 2050 when including supply chains, services, and ecosystem effects.
The uncertainty reflects genuine questions about how quickly the technology will mature, how fast costs will fall, and how readily industries will adopt humanoid workers. But the structural question is becoming clearer.
Both models will likely succeed, but in different ways. full-stack players like Tesla may capture 15-20% of global humanoid unit sales but 40-50% of total industry profits, dominating premium segments where customers value integrated experience. The open ecosystem will capture the majority of unit volume, enabling thousands of companies to build businesses on shared infrastructure.
The ecosystem's total value creation will dwarf the full-stack players because it will enable innovation at every layer: hardware manufacturers, component suppliers, software developers, training data providers, maintenance services, and application-specific innovators. The robots themselves are infrastructure. The real value lies in what they enable.
But how will this infrastructure actually operate? Who will deploy robots, maintain them, update their software, and ensure they work reliably day after day? The answer may look less like buying a car and more like subscribing to a mobile network. A new category of business is emerging: the humanoid network operator.