South Korean AI semiconductor startup DeepX is positioning its mass-produced DX-M1 chip as a low-power, cost-efficient alternative to Nvidia’s on-device GPUs as it seeks to gain ground in the emerging physical AI market.

The company said DX-M1 operates at an average power consumption of just 2 to 3 watts, delivering about 20 times higher power efficiency than Nvidia’s Jetson Orin for comparable workloads while costing roughly one-tenth as much. DeepX added that it has secured production yields exceeding 90 percent through collaboration with its foundry partner Samsung Electronics.

CEO Kim Lok-won said such advantages are critical as AI moves beyond data centers into real-world environments.

“When computing shifted from PCs to mobile, the key factors were power efficiency and heat control,” Kim said at a press briefing in Pangyo on Tuesday. “Even Nvidia, which dominates the AI chip market today, faces battery and thermal limitations in industrial and edge applications.”

Founded in 2018 by a former Apple chip designer, DeepX focuses on “physical AI,” referring to AI embedded in devices such as robots, autonomous vehicles and factory systems.

The company declared itself a “physical AI infrastructure company” and unveiled a three-stage strategy linking chips, hardware platforms and software ecosystems.

Under the model, DeepX will supply AI chips, while partners including Advantech, Dell and Raspberry Pi develop hardware platforms. Software partners such as Ultralytics and China’s Baidu will build AI applications on top.

DeepX is also working to lower switching barriers from GPU-based systems dominated by Nvidia. Its software platform DXNN enables models trained on GPUs to run on DeepX chips, while its DX-Newton interface improves compatibility with Nvidia’s Isaac ROS robotics framework.

Kim projected strong growth in the physical AI market, citing the expansion of robotics, smart factories and AI-powered devices.

“With the spread of AI devices, the physical AI market could grow more than threefold over the next five years,” he said. “DeepX aims to build the computing infrastructure layer for this market.”

The company also unveiled plans for its next-generation chip, DX-M2, which will adopt Samsung’s 2-nanometer process. The chip is designed to deliver up to 80 trillion operations per second at under 5 watts of power, with mass production targeted for 2027.

DeepX said the chip is aimed at enabling large-scale AI models to run on edge devices, bringing capabilities previously limited to data centers into real-world applications.

The company is also preparing for an initial public offering, with a domestic listing now seen as the most likely route.

Chief Financial Officer Cho Young said DeepX is targeting about $40 million in revenue this year, including $25 million from product sales, as it focuses on building a commercial track record ahead of the IPO.

By Jie Ye-eun (yeeun@heraldcorp.com)