South Korea's DEEPX Brings Low-Heat DX-M1 AI Chip to US, Claims It Runs 25 to 40 Degrees Cooler

South Korean on-device artificial intelligence chipmaker DEEPX introduced its low-heat AI chip, the DX-M1, in the United States on November 6. Chief executive Kim Nok-won said in an interview with CNBC on November 5 US time that the chip runs 25 to more than 40 degrees Celsius cooler than competing products; the company says it generates so little heat that butter placed on the chip does not melt. The DX-M1 is an in-house low-power neural processing unit (NPU) that consumes less than 5 watts while delivering high-performance inference with low latency, even during prolonged operation. (Cover photo: DEEPX product image.)

In mass production on Samsung 5nm for over a year; 88 orders in 10 countries

The DX-M1 has been in mass production on Samsung Electronics' 5-nanometre-class semiconductor process since August 2025. The Korea JoongAng Daily reported in April that the chip consumes an average of 2 to 3 watts, delivers about 20 times the power efficiency of Nvidia's on-device GPU, the Jetson Orin, at roughly one-tenth the price, and achieved a yield of more than 90 percent on Samsung's 5-nanometre process. DEEPX said that as of October 2026 it had secured 88 purchase orders across 10 countries totalling approximately $13.3 million, spanning robotics, smart factories, mobility and other industries.

Bound for Hyundai's first commercial robot; next chip targets large models

The DX-M1 will be installed in the first commercial robot scheduled for release this year by Hyundai Motor Group's Robotics Lab. The products DEEPX is collaborating on with Hyundai include the DAL-e autonomous mobile robot delivery platform and the MobED mobility platform; DEEPX and the Hyundai-Kia Robotics Lab have worked together since 2023 on on-device AI controllers that let robots operate autonomously without relying on the cloud. DEEPX is also developing its next-generation chip, the DX-M2, on Samsung's 2-nanometre process, designed to run generative AI workloads including large language models at low power, with mass production targeted for the second half of 2027.