Aug. 20, 2025 –
On-device AI semiconductor company DEEPX (CEO Lokwon Kim) has announced that it has signed a 2nm process development agreement with Samsung Foundry and its design partner Gaonchips, for its next-generation generative AI chip, the DX-M2.
Under the contract, DEEPX becomes the first commercial customer of Samsung Foundry’s advanced 2nm Gate-All-Around (GAA) process for the AI chip and will commence silicon fabrication for the DX-M2 processor, which is designed for ultra-low-power generative AI inference on devices. Multi-Project Wafer (MPW) prototype production is planned for the first half of 2026, with mass production targeted for 2027.
This project is expected to play a pivotal role in expanding South Korea’s high-value semiconductor industry and accelerating the development of a global 2nm AI semiconductor ecosystem.
While mass production is targeted for 2027, securing Samsung Foundry's 2nm GAA process at this early stage was a strategic imperative for DEEPX. The company recognized that achieving true, practical on-device generative AI is nearly impossible with existing process nodes due to the immense computational demands and strict power limitations.
The superior power-to-performance ratio of the 2nm process—delivering approximately twice the power efficiency of the 5nm process used in DEEPX’s current DX-M1—is not just an improvement, but a fundamental requirement to overcome this technological barrier. This decision highlights DEEPX's commitment to delivering a future-proof solution that will define the next era of AI, rather than settling for incremental gains.
DEEPX’s existing 5nm-based vision AI chip, the DX-M1, is in mass production for use in robotics, smart cameras, and factory automation, and is being adopted by customers around the world. The DX-M2 extends the company’s portfolio to generative and multimodal AI, forming the core of its strategy to lead the transition towards a hyper-intelligent, AI-driven society.
Today, generative AI is largely powered by expensive, energy-intensive data centers. The DX-M2 chip is designed to replace this traffic with an ultra-low-power, no-fee on-device AI solution that delivers tens of times higher power efficiency than data centers while significantly reducing carbon emissions.
The chip is capable of real-time inference up to 20 billion parameter models at a rate of 20 to 30 tokens per second (TPS), all while consuming 5 watts or less of power. This represents a monumental leap in power efficiency, especially when global competitors are currently running smaller 10-billion-parameter models at around 10 TPS with a power consumption of 10 to 20 watts. This allows expert-level generative AI models to run independently on devices such as robots, home appliances, and laptops, even in environments with tight power and thermal constraints.
DX-M2 is set to deliver the world’s first practical on-device solution capable of real-time processing of DeepSeek and LLaMA 4 Scout—20 billion-parameter models that achieve 100 billion-class near-AGI performance through a mixture of experts (MoE) architecture—as well as other advanced generative AI models in real time.