DeepX claims 100B-parameter inference under 5W
The gist
South Korean chip startup DeepX taped out its DX-M2 AI accelerator, targeting 100B-parameter models under 5W.
The claim, if it holds, would shift large-model inference from data centers to devices.
How it unfolded
- Aug 2025DeepX's first chip, the DX-M1, entered mass production on Samsung Electronics' 5-nanometer process.
- Jan 9, 2026At a CES 2026 media briefing, DeepX unveiled its DX-M2 roadmap, targeting on-device inference for 20B–100B parameter models within sub-5W.
- Jan 14, 2026DeepX announced the DX-M2 at CES 2026, designed to run models of up to 100 billion parameters at under 5 watts.
Key terms
- taped out
- The final stage of chip design when the finished layout is sent to a factory for manufacturing.
- AI accelerator
- Specialized hardware designed to speed up artificial intelligence and machine learning workloads.
- fabless
- Describes a semiconductor company that designs and sells chips while outsourcing their actual fabrication to specialized foundries.
Source: Wikipedia