Heat Sink Solutions：Market Size and Growth Trends

Computing Power Enhancement and Heat Dissipation Bottlenecks

A chip’s TDP (Thermal Design Power) of 350W is generally regarded as the dividing line between air cooling and liquid cooling. The high demand for AI computing power has accelerated chip iteration, with power consumption significantly increasing alongside performance upgrades. For instance, NVIDIA’s B200 chip released at GTC2024 has a heat output of up to 1,200W under full load, and the DGX B200 8-card server consumes nearly 15kW of power, prompting the launch of the GB200 NVL72 liquid-cooled rack system. This trend has not only driven up the power density of single cabinets in data centers but also highlighted the growing limitations of traditional cooling methods, making liquid cooling a critical solution to overcome heat dissipation bottlenecks.

Policy Drivers and Energy Efficiency Requirements

In July 2024, four Chinese government departments including the National Development and Reform Commission issued the Special Action Plan for Green and Low-Carbon Development of Data Centers, setting clear energy efficiency targets: by the end of 2025, the average PUE of data centers nationwide will be reduced to below 1.5, while the PUE of newly built, expanded, or renovated large and super-large data centers will be controlled within 1.25, and data center projects in national hub nodes must have a PUE no higher than 1.2. The policy also emphasizes “promoting efficient cooling and heat dissipation technologies like liquid cooling according to local conditions” and requires newly built or renovated data centers to adopt GPU unit computing power efficiency standards, providing policy support for the large-scale application of liquid cooling.

Domestic Chip Demands and Liquid Cooling Advantages

Affected by process technology and yield rates, the energy efficiency of domestic AI chips still has room for improvement, placing higher demands on equipment heat dissipation capabilities. Liquid cooling technology, with its core advantages of low energy consumption, high heat dissipation efficiency, low noise, and low TCO (Total Cost of Ownership), not only meets the heat dissipation needs of high-density computing scenarios but also aligns with the goals of building green data centers. With the combined momentum of policy support and market demand, liquid cooling technology is expected to enter a rapid growth phase amid the wave of AI computing power upgrades.

The AI Chip Arms Race Will Continuously Drive Product Upgrades, with Medium-to-Long-Term Supply Likely to Diversify

In the short to medium term, the development and competition of AI models will continue to boost AI chip shipments and specifications. In the long term, AI chip demand will focus more on the input-output ratio and total cost of ownership (TCO), with the core expected to shift from large AI model training to vertical AI model training and AI inference. According to a Yole report, the penetration rate of AI servers (including GPUs and other accelerators) is expected to increase from nearly 10% in 2023 to over 18% by 2028, of which approximately 70%-75% will be GPU servers.

Scale of Accelerator Servers and Proportion of GPU Servers (Left: Millions of Units; Right: %)

Moore’s Law Slows, with Chip Computing Power and Power Consumption Both Rising

IDC estimates that China’s intelligent computing power scale was approximately 260 EFLOPS (FP16) in 2022 and will increase to 1,117 EFLOPS by 2027, with a compound annual growth rate (CAGR) of 34% from 2022 to 2027. The high demand for computing power has driven the accelerated iteration of AI chips, with significant increases in power consumption alongside performance upgrades. From 2016 to 2022, the average power consumption of CPUs rose from 100–130W to 300–400W, while that of GPUs/NPUs increased from 250W to 500W. NVIDIA’s single H100 has a maximum TDP (Thermal Design Power) of 700W, and the newly released B200 using the Blackwell architecture has a power consumption of 1,000W. The GB200 solution, composed of two B200 GPUs and one Grace CPU, has a power consumption as high as 2,700W, shifting the heat dissipation approach from air cooling to liquid cooling. A TDP of 350W is generally regarded as the dividing line between air cooling and liquid cooling, and it is expected that most of the latest generation of processors will have a TDP exceeding 400W within the next three years, surpassing the heat dissipation capacity of air cooling.

With the continuous advancement of chip technology, while chip performance has increasingly improved, their power consumption has also increased, leading to a sustained growth trend in the chip heat dissipation market. Relevant data shows that the global data center thermal management market size was $16.56 billion in 2024 and is expected to grow to $34.51 billion by 2029, with a compound annual growth rate (CAGR) of 15.8% from 2024 to 2029. In terms of technology penetration, Omdia estimates that the combined market size of air cooling and liquid cooling for data centers was $7.67 billion in 2023, with liquid cooling accounting for approximately 17% of the market. In the consumer electronics sector, global smartphone shipments reached 1.17 billion units, tablet shipments 130 million units, and computer shipments 250 million units in 2023, driving the global consumer electronics thermal management market to an estimated $30.9 billion.

By 2030, the global data center thermal management market is expected to reach even higher levels, with significant growth also anticipated in emerging chip heat dissipation markets. Technologies such as diamond cooling and vapor chamber (VC) cooling are poised to capture larger market shares. By 2030, the global chip-level heat dissipation market is projected to reach $40 billion, with a compound annual growth rate of 31.4%.