When the AI ​​chip "fever", who is guarding the hot machines with their noses?

The Wall Street Journal recently profiled an underrepresented group of people in U.S. data centers.

They don’t write code or stare at data dashboards. I walk tens of thousands of steps in a computer room of tens of thousands of square meters every day. My tasks include listening to whether the sound of the fan is correct, feeling whether the cabinet is hot, and getting closer to smell if there is a burning smell in the air.

Why do you smell burnt? Because AI servers are filled with GPUs, high-power power supplies, and densely packed electronic devices, they run at full capacity for a long time and generate astonishing amounts of heat. Cables are aging, power modules are overheating, and a slight burnt smell will first emerge from the plastic insulation layer - workers often detect problems earlier than the alarm system by relying on their noses.

They are the operation and maintenance workers of the data center and the most "silent" line of defense in the AI ​​era.

The AI ​​in your mobile phone is backed by hot chips.

When you ask a question to ChatGPT on your phone, the question is sent to a data center. There were endless rows of metal cabinets, lights flashing on and off, and almost no one in sight.

But inside these cabinets, a silent heat war is taking place.

According to data from the Ministry of Industry and Information Technology, the current power consumption of a single AI GPU chip has reached 1,400W, and is expected to exceed 2,000W in 2027, with a local heat flow density exceeding 1,000W/cm². What concept? A chip the size of a fingernail has a local heat flux density of over 300 W/cm², several times that of a nuclear reactor core. The limit of traditional air cooling is only about 100 W/cm², which is already beyond its capabilities.

Data center power density is also soaring. In 2000, the power of a single rack was only 2 kilowatts. In 2025, the AI ​​rack has reached 140 kilowatts, and the next generation target is 250 kilowatts - equivalent to more than 100 electric kettles squeezed into one square meter of space to boil water at the same time. Traditional air cooling has reached its physical limit at 25 kWh, and will face "thermal runaway" if it goes higher.

The operation and maintenance workers are the people patrolling next to this "big boiler".

Uptime Institute"s 2025 Global Data Center Survey shows that nearly two-thirds of operators report difficulty retaining employees or finding qualified candidates. At the same time, the cooling requirements brought by AI technology are making traditional operation and maintenance strategies more complex.

There are not enough manpower, and the machines are getting hotter and hotter. This is the reality of data center operation and maintenance in the AI ​​era.

Where does the heat come from?

There is a clear timeline for the rapid increase in GPU power consumption: in 2017, the NVIDIA V100 chip power was about 300 watts; in 2024, the Blackwell chip soared to 1,200 watts - equivalent to a household electric kettle; the next generation Rubin architecture is expected to exceed 2,000 watts.

The core bottleneck of traditional heat dissipation solutions is not that the 'coolant is not cold enough', but that the heat diffuses from the surface of the chip to the surroundings of the cold plate fast enough . The thermal conductivity of pure copper is about 380-400 W/(m·K), and that of aluminum is only about 200 W/(m·K). When the heat flux density reaches 1000 W/cm², the heat will produce serious 'heat accumulation' in the central area of ​​the chip, and the rate of heat diffusion of the substrate itself cannot keep up with the rate of heat generation by the heat source.

Increasing the coolant flow rate will not solve this problem because the bottleneck is heat conduction within the solid material, not fluid convection.

An opinion article published by Professor Yang Ronggui of Peking University and Professor Li Xiaobo of Huazhong University of Science and Technology in Science Bulletin pointed out that thermal management is evolving from a "remedial measure" after the chip design is completed to a core link of "boundary setting" at the front end - the peak performance of the chip is no longer determined by the number of transistors, but by the heat dissipation capacity.

If heat can be spread instantly at the source

If the thermal conductivity of the heat sink material itself is increased by an order of magnitude, the heat will be quickly spread out before reaching the coolant - and the workers may smell less burnt smell when inspecting.

This is exactly what Debaishun Blue Diamond Technology is doing.

Debaishun focuses on the R&D and production of diamond heat sink sheets. Its core product - 12-inch wafer-level conductive diamond heat sink , is designed for extreme thermal management scenarios such as AI computing chips, high-power semiconductors, laser devices, and low-altitude satellite communications.

Key performance indicators:

Thermal conductivity >1500W/(m·K) : Calculated using a multi-layer parallel thermal resistance model of 2μm thick blue diamond film and silicon substrate, it is at the leading level in the industry under the same conductivity conditions.

Direct use without polishing : Randomly sent samples AFM test shows that the surface roughness Ra of the sample without CMP polishing treatment reaches 42.1nm and 45.5nm, and the as-grown surface meets the packaging-level roughness requirements, eliminating the CMP polishing process required by traditional diamond, and comprehensively reducing costs by more than 30%.

12-inch wafer level : One of the few conductive diamonds in China that has achieved 12-inch CVD heterogeneous nucleation and continuous growth, non-mosaic splicing, no grain boundary thermal resistance, and meets the rigid requirements of advanced packaging for large-size substrates.

At the technical level, Debaisun's homogeneous phase doping process achieves both electrical conductivity and thermal conductivity - doping boron to achieve controllable conductivity while maintaining ultra-high thermal conductivity satisfies the dual functions of chip grounding and heat dissipation. From MPCVD equipment research and development, heterogeneous nucleation process to growth processing, the entire chain is independently controllable, the equipment is localized, and the production capacity can be replicated within 4 months.

Every 1°C decrease is a victory for computing power

The selection of heat dissipation materials is no longer an engineering detail.

From the models trained by scientific researchers, to power supply, transmission networks, data centers, GPU servers, cooling systems, and finally to the AI ​​answers on your mobile phone - the industry chain has been stretched, and all the glory of AI has finally landed on a hot machine.

The diamond heat sinks of Debaishun Blue Diamond Technology are solving the most difficult link in this chain from the material level: allowing the heat to be quickly taken away the moment it is generated, instead of waiting until it turns into a burnt smell before it is discovered.

Because every drop in temperature means more computing power can be released.