Taipei—Korea IT Times Reporting Team
At COMPUTEX 2026, Physical AI was no longer presented as a distant concept. At the NVIDIA Inception startup pavilion inside the InnoVEX Zone on the fourth floor of TaiNEX Hall 2, the shift was visible in a more practical form: AI startups were not only demonstrating technology, but also showing how quickly it could be turned into industrial products.
The pavilion was not defined by oversized robots or theatrical hardware displays. Its real message was commercialization. Startups from across Asia were using NVIDIA’s platforms to shorten the path from prototype to deployment in robotics, edge AI, data infrastructure, bio-AI, and industrial automation.
NVIDIA Inception is a global program that supports more than 19,000 AI startups. At COMPUTEX 2026, about 30 selected companies were brought together inside the joint pavilion. Korean, Taiwanese, Singaporean, and Japanese startups used the space to present how their technologies could fit into real markets rather than remain at the demonstration stage.
What made the pavilion significant was not simply the number of promising startups on display. It showed how NVIDIA’s core technology stack — including GPUs, Omniverse, Isaac Sim, and Project GR00T — is beginning to shape how startups build products, validate systems, and enter global markets.
Korean Startups Stand Out in Robotics and AI Alignment
Korean startups drew attention in two areas: robotics and AI infrastructure.
AeiRobot placed its humanoid robot, ALICE4, at the front of the pavilion. The robot interpreted natural-language commands and performed unstructured pick-and-place tasks, attracting steady interest from visitors. The demonstration was not just about movement. It showed how a robot trained in a simulated environment could operate in a physical setting.
An engineer at the booth said the company had used NVIDIA’s GR00T foundation model and Isaac Sim-based simulation to reduce errors before real-world deployment. ALICE4’s selection for the NVIDIA Award at the InnoVEX pitch contest reflected that practical validation.
Coxwave took a different position. Instead of showcasing a robot, it focused on the operational layer behind AI models. Its Coxwave Align solution presented tools for alignment, performance analysis, and infrastructure optimization for generative AI and multimodal models.
As AI models move into industrial environments, the central issue is no longer model performance alone. Controllability, traceability, and operational reliability are becoming critical. Coxwave’s exhibit suggested that tools for understanding model behavior, reducing errors, and improving developer productivity are becoming core infrastructure for AI commercialization.
Taiwanese Startups Target Edge AI for Industrial Sites
Taiwanese companies, representing the host country, leaned heavily on the island’s manufacturing and semiconductor supply-chain strengths. Their focus was not on cloud-heavy AI systems, but on practical edge AI and domain-specific models designed to operate directly in factories, rail systems, waste-treatment facilities, and healthcare environments.
AIRECO introduced an AI solution for automating waste-treatment facilities. The company emphasized the potential to reduce human exposure in high-risk work environments while supporting more efficient and environmentally responsible operations.
POD Innovation presented a forward-collision prevention system for light rail transit. The system uses edge vision AI to detect risk factors in urban mobility environments in real time.
APMIC introduced S1 Distillation, a model distillation and lightweighting technology. The solution is designed to help enterprises and manufacturing sites operate private small language models without sending sensitive data to external cloud systems.
Avalanche Computing addressed one of the biggest barriers to enterprise adoption of generative AI: data leakage. Its AIthena platform presented a structure for using GenAI while protecting internal corporate data.
Morale AI showcased a domain-specific LLM agent designed for industry-specialized knowledge. Rather than competing in the broad general-purpose model race, the company is targeting AI agents optimized for specific business and industrial problems.
MetAI Technology also stood out with MetGen, a generative AI solution for creating industrial 3D digital twins. The company’s presentation suggested that Taiwan is expanding beyond hardware manufacturing into industrial AI software integration.
Singapore and Japan Focus on Bio-AI and Data Infrastructure
Startups from Singapore and Japan targeted more specialized markets.
Singapore-based Nanyang Biologics introduced DTIGN, a graph neural network-based model designed to predict drug-target interactions. By accelerating the identification of potential drug candidates, the system points to how bio-AI is moving from data analysis into the core workflow of drug discovery.
Lynx Analytics presented LynxKite 2000, a platform for visualizing and tracking large-scale network structures. The company positioned the technology for business intelligence, finance, telecommunications, and other sectors that rely on interpreting complex data relationships.
Japan’s APTO introduced a crowdsourced data collection and high-quality labeling platform for AI training. As generative AI and Physical AI advance, data quality is becoming one of the most important constraints on performance. APTO is positioning itself as an infrastructure provider for that bottleneck.
The New Link Between Platforms and Applications
The NVIDIA Inception pavilion showed a clear shift in startup strategy. Companies are no longer relying only on the claim that they can build large models. Instead, they are using NVIDIA’s technology stack to solve specific industrial problems faster.
Omniverse and Isaac Sim are becoming foundations for testing robots and manufacturing systems in virtual environments before deployment. Project GR00T is beginning to function as a shared foundation for humanoid robotics development. GPUs and AI software tools are helping startups bring early products to market with greater speed.
This structure creates both opportunity and dependency.
Startups that build on NVIDIA’s ecosystem gain development speed, technical credibility, and easier access to global partners. But as computing resources, simulation environments, and development tools concentrate around one platform, dependence on that platform also grows.
A global venture capital investor at the venue said the basis for judging AI startups has changed. “The question is no longer just how large the model is,” the investor said. “The question is how quickly the technology can be validated in an industrial setting. Companies inside the NVIDIA ecosystem clearly have an advantage in moving from prototype to commercialization.”
Reporter’s View
The NVIDIA Inception pavilion was both a startup showcase and a condensed view of NVIDIA’s platform strategy.
In earlier phases of AI competition, the question was often who could build the larger or more powerful model. At COMPUTEX 2026, the question had changed. The focus is now on which companies can solve industrial problems quickly, safely, and cost-effectively.
Korean startups showed promise in robotics and AI alignment. Taiwanese companies combined manufacturing supply chains with edge AI. Singaporean firms focused on bio-AI and data analytics infrastructure, while Japanese companies targeted the foundational market for AI training data quality.
Behind all of these efforts, NVIDIA provided the common infrastructure. The startups were addressing different industrial problems, but the computing power, simulation tools, and development environments behind them were increasingly converging around a single platform ecosystem.
What the NVIDIA Inception pavilion revealed was another side of the Physical AI transition. The most important development was not only that robots were moving. It was that the startup ecosystem and platform power behind those robots were already being reorganized.
Editorial Contact: Korea IT Times Technology Team (monica@koreaittimes.com | info@koreaittimes.com)
