DX-M2 brings large generative AI out of the data center and into robots, vehicles, AI PCs, and other intelligent machines. Today, models above 10 billion parameters usually need GPU-class computing or cloud infrastructure, which adds power, heat, latency, connectivity, and recurring cost. DX-M2 is designed to change that. Built on a 2 nm process, its 5W-class accelerator is designed for 80 TOPS and supports up to 96 GB of 8-channel LPDDR5X memory within an approximately 10W module envelope. That combination is designed to let 10B+ language and multimodal models run locally. Hardware-native mixed precision and the DXNN software stack further reduce model size and data movement. The result is simple: more capable AI can live inside the machine that needs it, instead of somewhere in the cloud.