DeepX vs. Nvidia: Korean Chip Startup Claims 20x Power Efficiency

South Korea's DeepX is betting that the next wave of AI will not run in data centers. The startup, founded in 2018 by former Apple chip designer Kim Lok-won, makes processors for robots, autonomous vehicles, and factory systems. Its pitch: the DX-M1 chip delivers 20x better power efficiency than Nvidia's Jetson Orin at roughly one-tenth the cost.

Average power draw: 2 to 3 watts.

The Details

DeepX calls its target market "Physical AI," machines that process sensor data and make decisions in the real world rather than on a cloud server. The DX-M1 is manufactured at Samsung's foundry with yields above 90%.

The software layer matters as much as the silicon. DXNN, DeepX's proprietary framework, lets developers take models trained on Nvidia GPUs and run them directly on DeepX chips. DX-Newton, the company's hardware interface, is compatible with Nvidia's Isaac ROS ecosystem. In practice, that means customers can switch from Nvidia to DeepX without rewriting their software stack.

Partners: Advantech, Dell, and Raspberry Pi on the hardware side. Ultralytics and Baidu on software.

Revenue target: roughly $40 million in 2026, with $25 million from product sales. DeepX is planning a domestic Korean IPO, though no timeline has been announced.

Next-gen DX-M2

The successor chip, DX-M2, will use Samsung's 2nm process and deliver up to 80 TOPS at under 5 watts. Mass production is scheduled for 2027.

The cost equation for edge AI

Nvidia dominates AI training and cloud inference. But edge devices, robots on factory floors, drones, delivery vehicles, need chips that run on battery power in dusty environments, not $10,000 modules designed for server racks. DeepX is one of several startups targeting this gap, but few have the combination of Samsung foundry access, Nvidia software compatibility, and hardware partners already shipping.

Kim puts the opportunity in simple terms: "The physical AI market could grow more than threefold over the next five years." At 2 to 3 watts per chip, the math favors whoever can deliver usable AI performance at the lowest power budget. That race is still early.

Sources: The Investor, DeepX