South Korean chip startup DEEPX showed a block of butter staying solid on its DX-M1 AI chip while the same butter sizzled on competing silicon.
In the side-by-side demo, each chip ran the same AI workload. CEO Kim Nok-won said "every chip melted the butter except the last chip," meaning the DX-M1. He put the temperature gap at 25 to more than 40 degrees Celsius. The rival chips were not named.
The DX-M1 is an NPU, or neural processing unit, a chip built only for AI math. It has been made on Samsung's 5nm process since August 2025 and has won about $13.3M in orders. The next chip, the DX-M2, moves to Samsung's 2nm process, with mass production targeted for the second half of 2027.
The demo is the company's own, and it leaves out the number that matters most: how fast each chip finished the job. Any processor runs cooler if you cut its power and clock speed, which is why PC enthusiasts undervolt their GPUs. Butter measures heat, not efficiency. Performance per watt would be the useful comparison, and it is missing.
The bigger mismatch is the arena. The DX-M1 targets edge computing and robotics, meaning AI that runs on a device instead of in a server farm. It does not compete with the Nvidia Blackwell and Rubin racks that motivated the demo's framing, where heat, power and cooling have fueled a growing backlash against data centers.
The more relevant work happens elsewhere. SK hynix is developing iHBM, a memory design whose "i" stands for Integrated Cooling Element, according to one of its executives. CoolIT is pushing water cooling closer to the silicon. Both go after heat at data center scale. Whether a chip designed for low power can ever stretch to that scale is the question the butter cannot answer.