For the past few years, embodied AI competition has focused on increasingly capable robots and foundation models with better perception, reasoning, planning and control. Those advances still matter, but the competitive question is shifting. The next stage will depend less on impressive demonstrations and more on how quickly companies can turn models into reliable physical products, learn from deployment and iterate.
China offers a clear view of that transition. More than 70 embodied AI training grounds were operating by the end of June 2026, with another 46 under construction or planning, according to a report cited by TechNode. Industrial manufacturing appeared in 86 percent of them. The significance is not simply the scale of investment, but the push to expose robots to the variability of real-world environments, where many of the hardest commercialization problems begin.
A good AI model is not yet a good machine
It has become increasingly possible to demonstrate sophisticated intelligence in a controlled environment. The harder task is keeping that intelligence useful when it encounters different hardware, noisy surroundings, variable connectivity, unfamiliar users, changing physical conditions and tasks that do not unfold exactly as expected. Embodied AI combines several development cycles that traditionally moved at very different speeds.
AI models can now change within weeks or months. Hardware design moves much more slowly. Firmware, sensors, connectivity, mechanical components, power constraints, safety requirements and manufacturing have their own timelines. If those layers are tightly coupled, every improvement to one part of the system can create integration work somewhere else.
A team may discover that a new model handles instructions substantially better, for example, but switching models should not require redesigning the voice interaction layer. A robot may need access to a new service or source of information, but adding that capability should not require redesigning the entire software stack. Flexibility is becoming more than an engineering preference. It is becoming part of the economics of embodied AI.
China’s robotics ecosystem is moving toward deployment
Several developments in China this year suggest that deployment is becoming a more explicit part of the industry’s competitive logic. NVIDIA, for example, has been hiring in Beijing, Shanghai and Shenzhen across embodied AI, simulation, deployment and solution architecture, with deployment teams focused on optimizing algorithms for humanoid robots and accelerating their use in real-world environments.
There is also a more fundamental commercialization test emerging. TechNode’s recent examination of Galbot’s industrial robotics strategy described cost, reliability, stability and large-scale deployment as continuing barriers to adoption. The important question was not whether a robot could complete a task once, but whether it could operate continuously in a production environment and generate measurable economic value.
An embodied model may perform extremely well on a benchmark while the finished device remains commercially difficult to deploy. Conversely, a system with slightly less impressive headline intelligence may create considerably more value if it can be integrated, updated, supported and scaled efficiently.
Iteration speed may become a competitive advantage
Software companies have long benefited from rapid iteration. Hardware companies historically have not enjoyed the same freedom. Embodied AI creates an opportunity to bring some of that software-style iteration into physical products, provided companies design for change. Models, external tools, connectivity and hardware interfaces should be treated as components that can evolve rather than permanent decisions made at the beginning of product development.
At IFA 2026 in Berlin, Agora demonstrated conversational AI across several connected-device form factors, including robots, AI companions, developer hardware and wearables. The broader point is that the intelligence and interaction layer can evolve without treating every new piece of hardware as an entirely new AI project.
This flexibility matters particularly for startups, which often make decisions about chipsets, models, connectivity, sensors and cloud services before they have enough real users to know which assumptions will hold. Locking those choices together increases the cost of being wrong.
The strongest architectures will therefore become more modular. Developers should be able to replace a model without rebuilding the device, introduce new tools without rewriting every integration and carry more of the interaction experience from one form factor to another. Teams capable of making those changes quickly will learn faster.
Deployment creates a new source of training data
Moving into the real world does more than validate a product. It creates feedback that is difficult to reproduce fully in simulation, from unexpected physical conditions to network degradation and user behavior.
China’s growing network of embodied AI training grounds reflects the importance of this learning loop. Deployment and model development are therefore becoming less distinct: real-world use generates information, that information improves the system and the improved system returns to deployment. Companies that can complete this cycle fastest may gain not only an operational advantage but a data advantage as well.
The winning robot may not start with the winning model
Commercialization introduces a different measure of progress. Can the machine work reliably outside a demonstration? Can it be updated as AI changes? Can developers learn from deployment quickly enough to improve the next version? Can the economics work at scale?
Those questions are less visually impressive than watching a robot perform a new feat, but they may ultimately determine which companies build sustainable businesses. The embodied AI race is not moving away from intelligence. Intelligence is becoming one layer of a much larger system. As strong models become more widely available, the ability to deploy them, learn from them and iterate around them may become one of the industry’s most important competitive advantages.
Lawrence Wu is the Head of Physical AI at Agora, a company specializing in real-time engagement technology. He holds a Master’s degree in Industrial Engineering and Engineering Management from National Tsing Hua University and a Bachelor’s degree from National Cheng Kung University.
Before joining Agora, Lawrence held sales and management positions in various technology companies, where he focused on driving business growth and forging strategic partnerships. His experience in these roles has equipped him with a deep understanding of market dynamics and customer needs, positioning him as a key contributor to Agora’s expansion into the IoT sector.
At Agora, Lawrence is responsible for leading digital transformation initiatives and driving innovation in real-time engagement solutions. His strategic vision and leadership have been instrumental in establishing Agora as a strong player in the real-time communication industry.
Editor’s note: This contributed article has been lightly edited for clarity and TNGlobal house style. The substance of the author’s contribution has been preserved.
Share your perspective: TNGlobal welcomes contributed insights and expert commentary from across Asia’s technology and innovation ecosystem. Submit a contribution for editorial consideration, or explore more conversations in our TNGlobal INSIDER and TNGlobal Q&A and Interviews archive.
Featured image: lhon karwan on Unsplash