DeepX Commits Over $100 Million to Samsung 2nm AI Chip

DeepX will spend more than $100 million, about 133 billion won, to build its next-generation AI chip on Samsung Electronics' 2-nanometer process. Bloomberg reported the plan on the 13th.

The South Korean chip designer has finished the design of the part, called DX-M2, and handed it to Samsung's foundry business. The budget covers performance verification and production engineering.

The timeline is set. DeepX targets first silicon by March 2027. Evaluation samples would reach customers in the following quarter. Initial production volumes are due in the fourth quarter of that year, with full-scale mass production starting in 2028.

What does the chip actually do?

DX-M2 is built to run large language models and vision language models at about 5 watts on Samsung's 2-nanometer node. That power envelope is the whole point. DeepX names robots, industrial equipment and AI-capable personal computers as the key applications.

Bloomberg framed the investment as a large-scale bet on running AI models on the devices themselves instead of relying on remote data centers, and the reasoning is straightforward economics: on-device AI can cut the energy, cost and time involved in sending data to the cloud and back.

A 5-watt chip that handles large models locally changes the arithmetic for battery-powered robots and industrial edge systems, where every watt counts and network links cannot be assumed, so the design choices point at a market the big accelerator vendors have largely left alone.

What does this change for Samsung's foundry?

The deal gives Samsung's 2-nanometer process another committed customer at a scale above $100 million. Foundry revenue at that node depends on design wins like this one landing and staying on schedule through 2027 and 2028.

A second front: on-premise servers

DeepX announced a separate effort the same day. It will work with Intel Korea and KTNF, a South Korean server maker, on on-premise inference servers that run AI inside corporate networks.

The division of labor is clear:

• DeepX: neural processing units for AI inference and model optimization technology
• Intel Korea: central processing units, extended processing unit platforms and related software
• KTNF: linking its server design, manufacturing and system integration capabilities to the effort

Why two tracks at once?

The moves fit one strategy. DX-M2 pushes AI compute onto small devices. The Intel Korea and KTNF partnership pushes it into private corporate infrastructure. Both aim at customers who want inference without depending on a cloud provider.

The next hard checkpoint is March 2027. If DeepX secures first silicon on schedule, evaluation customers get samples that quarter and the path to 2028 mass production stays open. Any slip pushes the whole chain back, and the $100 million budget buys no protection against calendar risk at an unproven node.