DeepX Unveils 2nd Generation Chip Roadmap for Sub-5W Ultra-Low Power LLM Inference at CES 2026

On-device AI semiconductor startup DeepX unveiled the development roadmap for its second-generation AI chip, ‘DX-M2,’ at CES 2026. On the 6th (local time), DeepX held a global media briefing in Las Vegas, announcing the DX-M2 development plan aimed at running large language model (LLM) inference from 20 billion (20B) parameters up to 100 billion (100B) parameters in an MOE structure, all under ultra-low power consumption of less than 5W. The goal is to realize 'Generative Physical AI,' which expands generative AI from data centers to edge devices. Choi Young-ho, Vice President of DeepX, stated, "Operating a 100B LLM at 5 watts signifies a change in the prerequisites of AI infrastructure, going beyond just improvements in high-performance computing," adding, "Generative AI will transition into a capability that exists anywhere, outside of data centers." DeepX estimates that if inference tasks are distributed across billions of edge devices, data center traffic could be reduced by more than 80%, providing an alternative to alleviate excessive power consumption and infrastructure bottlenecks. Through its first-generation chip, DX-M1, DeepX has secured over 50 mass production projects and is currently commercializing in sectors like robotics, defense, and smart home appliances. The company plans to concentrate the know-how in power management, heat control, and software stack optimization accumulated from the DX-M1 into the DX-M2. DeepX also presented three key execution pillars for market penetration. They are collaborating with Ultralytics and Baidu to establish open-source Physical AI development standards and expanding their global distribution network with partners like Avnet, DigiKey, and WPG. They also introduced a Dev-kit called ‘AI Player’ to simplify testing of edge AI applications. DeepX stated, "If the DX-M1 proved the mass production potential of vision AI, the DX-M2 will be a turning point that extends generative AI into the physical world," emphasizing, "The goal is to liberate AI from the constraints of electricity and server costs."