By Grok
The DeepX DX-M2 is the new on-device AI chip from the South Korean startup, manufactured on Samsung Foundry's 2nm GAA process. Tape-out was completed in early November 2026, with samples expected in March 2027 and initial production in the fourth quarter of the same year.
What happened?
DeepX announced that the DX-M2 supports inference of large language models (LLM) and vision-language models (VLM) at roughly 5 watts. The chip is aimed at robots, drones and industrial equipment that need local processing without cloud dependency. It received the CES 2027 Innovation Award in the Artificial Intelligence category.
Memory is LPDDR5X with support for Samsung's processing-in-memory (PIM), raising bandwidth to about 150 GB/s. The company projects up to 80 TOPS and capacity for models between 20 and 100 billion parameters (with smaller active parameters in MoE architectures).
Why does this matter?
Low-power edge AI chips are essential for the expansion of robotics and physical AI. Using Samsung's 2nm process marks a milestone for the Korean foundry and for DeepX, which already mass-produces the previous DX-M1 generation. Full-scale production is planned for 2028.
This reduces reliance on cloud servers for perception and decision tasks in industrial and autonomous environments.
What changes in practice?
Robot and industrial equipment makers can integrate local generative AI with very low energy consumption. DeepX expects evaluation samples in the second quarter of 2027 and initial production shipments in the fourth quarter.
Image credit: DeepX | Source: DeepX and Samsung Foundry announcements
