DEEPX DX-M1 NPU Runs Cool Enough to Keep a Butter Heatsink Solid

A knob of butter sitting unmelted on a running AI processor is the strongest claim DEEPX can make, and the South Korean startup made it on camera. The company told CNBC its new DX-M1 neural processing unit will sit at the heart of upcoming Hyundai robots, and it staged a dairy-based thermal test to prove the chip's efficiency credentials.

In the side-by-side demonstration, unnamed rival AI chips caused a block of butter to sizzle within seconds. The butter placed atop the DX-M1 remained solid. CEO Kim Nok-won claimed a temperature gap of 25 to more than 40 degrees Celsius against competitors running what he described as the same workload. "Every chip melted the butter except the last chip," Kim said.

What does the butter test actually prove?

Less than it appears. The comparison never disclosed how fast each processor completed the workload, and any chip can run cool if power and clocks are scaled back. PC enthusiasts undervolt and underclock their GPUs routinely for exactly that trade. The rivals also went unnamed. It is a memorable demo; it is not a controlled benchmark.

The commercial facts are firmer. According to The Elec, the DX-M1 NPU has been produced on Samsung's 5nm process since August 2025, and that single piece of silicon has won roughly $13.3 million in orders. DEEPX plans a successor, the DX-M2, on Samsung's 2nm node, with mass production targeted for the second half of 2027.

Where does this fit in the AI silicon market?

The context matters more than the stunt. CNBC's report frames the DX-M1 against growing unease over AI data center proliferation, which is reshaping business decisions, environmental debates and politics; hundreds of protesters have risked jail over data center construction. The core problem, per the report, is hot-running processors and the extensive power and cooling apparatus they demand.

DEEPX's answer, in the company's own words, is to "make the AI chips themselves run cooler." But the DX-M1 and DX-M2 target edge computing and robotics, not the racks of power-hungry Nvidia Blackwell and Rubin GPUs that dominate the CNBC feature. The butter test is, in effect, a comparison across categories. These NPUs were never going to melt anything in Nvidia's server territory.

Who else is attacking the heat problem?

The same CNBC report highlights two adjacent efforts aimed at the data center side of the equation:

• SK hynix is developing iHBM, a next-generation memory line where the "i" stands for Integrated Cooling Element, according to an executive who spoke to CNBC.
• CoolIT is pursuing a similar direction, moving water cooling closer to the heat-generating silicon itself.

Both approaches treat the symptom at the package level rather than reducing the heat the silicon produces in the first place. DEEPX's pitch is the opposite bet: fix the source.

The long-term implication, if cooler-running custom accelerators mature, is that the hundreds or thousands of hot GPUs fitted with bulky liquid cooling inside AI data centers could eventually be displaced. That shift would do more than cut power bills; it could cool the political backlash now forming around data center construction.

The verdict

DEEPX has a real product, a real process node, real orders and a real design win pipeline that includes Hyundai robots. The butter demo, by contrast, proves thermodynamics rather than competitiveness. Until the DX-M2 reaches mass production in the second half of 2027 and faces named rivals at matched performance, the dairy benchmark should be enjoyed as theater, not evidence.