South Korean AI chip startup DEEPX has demonstrated the cool-running capabilities of its DX-M1 processor by placing a knob of butter on the chip, which remained solid while competitor chips caused the butter to sizzle within seconds, in a side-by-side test reported by CNBC.
The demo, featured by CNBC, shows the DX-M1 processing the same AI workload as unnamed rival chips, according to DEEPX CEO Kim Nok-won, who claimed a temperature difference of 25 to more than 40 degrees Celsius compared with competitors. Kim said every chip melted the butter except the last one, the DX-M1.
DEEPX says its approach to addressing concerns over heat generation and power use in data centers is to make the AI chips themselves run cooler. The company has produced the DX-M1 on Samsung's 5nm process since August 2025, according to reports from The Elec, and has won approximately $13.3 million in orders. The firm plans to move forward with the next-generation DX-M2 AI chip based on Samsung 2nm, aiming for mass production in the second half of 2027.
Tom's Hardware noted gaps in the comparison information, particularly how fast each processor handled the test workload, as any chip can run cool if power and clocks are scaled back. According to Tom's Hardware's analysis, these neural processing units look set for edge computing and robotics applications rather than competing in the same space as power-intensive data center GPUs such as Nvidia's Blackwell and Ruben GPUs.
The CNBC report also highlighted SK hynix's work on iHBM, a next-generation, low-power, cool-running technology whose i an executive told CNBC stands for Integrated Cooling Element. CoolIT is also trying a similar approach by bringing water cooling closer to the heat-generating silicon.
Analysis
Why This Matters
- Rising energy consumption and heat from AI data centers are drawing environmental and political backlash globally, with protests and regulatory scrutiny increasing.
- DEEPX's demonstration underscores the potential for more efficient chips to reduce data center costs and resource use, though its current products target edge computing rather than the largest data centers.
- The company's next-gen DX-M2 chip, aimed at 2027, may broaden its addressable market if thermal efficiency scales.
Background
DEEPX is a South Korean semiconductor startup focused on low-power AI accelerators for on-device and edge applications. Data centers, which rely heavily on high-performance GPUs from companies like Nvidia, face growing criticism for their electricity and water demands, driving interest in more efficient custom chips.
Key Perspectives
DEEPX CEO Kim Nok-won: Claims the DX-M1 runs 25 to over 40 degrees Celsius cooler than competitor chips while processing the same AI workload, demonstrated by the butter test.
Tom's Hardware analysis: Notes the lack of comparison on processing speed and that the DX-M1 is not designed to compete with data center GPUs, making the test a red herring.
CNBC report: Highlights broader industry efforts such as SK hynix's integrated cooling memory and CoolIT's water-cooling approach to manage heat.
What to Watch
- Whether DEEPX provides detailed performance benchmarks alongside thermal results for the DX-M1.
- The progress of the DX-M2 chip on Samsung 2nm, with mass production targeted for H2 2027.
- How data center operators respond to cooling innovations, including integrated and liquid cooling solutions from SK hynix and CoolIT.
Checked
11 claims in this story checked against 1 source: 10 supported.
Show each claim
- SupportedDEEPX used butter to demonstrate that its AI chip runs cooler.Latest from Tom's Hardware: “South Korean chip startup DEEPX, recently featured by CNBC, thinks its AI chip designs are the answer. DEEPX demonstrated its DX-M1 processor’s cool-running capabilities in a bizarre new dairy-based thermal benchmark.”
- Partly supportedThe DX-M1 stays 25 to more than 40°C cooler than rivals, according to the CEO.Latest from Tom's Hardware: “DEEPX CEO Kim Nok-won claims a temperature difference of 25 to more than 40 degrees Celsius compared with competitors.”
- SupportedIn the demo, rival chips made a knob of butter sizzle within seconds while the butter on the DEEPX chip remained solid.Latest from Tom's Hardware: “While rival chips caused a knob of butter to sizzle in seconds, atop the DEEPX solution, the butter block remains solid.”
- SupportedThe side-by-side demo compared the DX-M1 with unnamed rival AI chips processing the same workload, according to Kim.Latest from Tom's Hardware: “In a side-by-side test against unnamed rival AI chips, you can see the DX-M1 NPU is clearly cooler. Kim said the side-by-side chips processed “the same AI workload…”
- SupportedKim said every chip melted the butter except the last chip, the DX-M1.Latest from Tom's Hardware: “Kim said the side-by-side chips processed “the same AI workload… and every chip melted the butter except the last chip [DX-M1].””
- SupportedThe DX-M1 has been produced on Samsung’s 5nm process since August 2025, according to The Elec reports.Latest from Tom's Hardware: “DEEPX’s DX-M1 NPU has been produced on Samsung 5nm since August 2025, according to The Elec reports.”
- SupportedDEEPX has won approximately $13.3 million in orders.Latest from Tom's Hardware: “It has won approximately $13.3M worth of orders thanks to this sliver of silicon.”
- SupportedDEEPX plans to move forward with the DX-M2, based on Samsung 2nm, aiming for mass production in the second half of 2027.Latest from Tom's Hardware: “The company is planning to move forward with the next-gen DX-M2 AI chip, which is based on Samsung 2nm, with its mass production aiming for H2 2027.”
- SupportedTom’s Hardware noted that the comparison did not say how quickly each processor handled the workload, and that scaling back power and clocks can make a chip run cool.Latest from Tom's Hardware: “Most notable was the question of how fast each processor tackles the test workload. Anyone can run a chip cool if the power and clocks are scaled back.”
- SupportedTom’s Hardware said these NPUs look suited to edge computing and robotics rather than competing in the same space as Nvidia’s Blackwell and Ruben GPUs.Latest from Tom's Hardware: “In conclusion, these NPUs look set for edge computing and robotics applications, so they aren’t really competing in the same space as the racks of power-gulping Nvidia Blackwell and Ruben GPUs featured in the CNBC news feature.”
- SupportedThe CNBC report highlighted SK hynix’s next-generation, low-power, cool-running iHBM technology, whose “i” stands for “Integrated Cooling Element,” according to an executive speaking to CNBC.Latest from Tom's Hardware: “The CNBC report also highlights SK hynix’s work on iHBM, a next-gen low-power consumption, and cool-running. The ‘i’ stands for ‘Integrated Cooling Element,’ according to an exec talking to CNBC.”
