The upcoming launch of NVIDIA’s Blackwell platform in Q4 2024 is poised to significantly accelerate the adoption of liquid cooling solutions in AI servers.
The penetration of liquid cooling systems is projected to increase from 10 percent in 2024 to over 20 percent by 2025, driven by rising environmental, social, and governance (ESG) awareness and the rapid deployment of AI servers by cloud service providers (CSPs), according to TrendForce.
The TrendForce report said NVIDIA is the leading supplier in the global AI server market, maintaining a nearly 90 percent market share in the GPU AI server segment for 2024, while AMD trails with just 8 percent.
The high energy demands of the new Blackwell platform, particularly the GB200 full-rack solution, are pushing the need for more efficient cooling solutions, prompting the shift from traditional air cooling to liquid-based systems. However, server original design manufacturers (ODMs) still face challenges such as potential leakage and cooling inefficiencies as they adopt these new technologies.
NVIDIA says Blackwell-architecture GPUs pack 208 billion transistors and are manufactured using a custom-built TSMC 4NP process. All Blackwell products feature two reticle-limited dies connected by a 10 terabytes per second (TB/s) chip-to-chip interconnect in a unified single GPU.
TrendForce forecasts that Blackwell could capture over 80 percent of the high-end GPU market by 2025, opening opportunities for power supply and cooling solution manufacturers to enter the liquid cooling sector.
Taiwanese suppliers are expected to supply quick disconnect components by the first half of 2024, with Google taking the lead in deploying liquid cooling for its AI hardware, particularly for TPU chips.
Major CSPs like Google, AWS, and Microsoft have accelerated AI server deployment, using NVIDIA GPUs and custom ASICs. As Blackwell’s GB200 racks — with thermal design power (TDP) reaching 140 kW — require efficient liquid cooling, Liquid-to-Air (L2A) technology is expected to become the dominant approach. Lower-density server architectures, like NVIDIA’s HGX and MGX platforms, will continue to rely on air cooling.
The industry is preparing for a new competitive landscape, with companies like BOYD, Cooler Master, and Vertiv leading the development of critical liquid cooling components. Taiwanese firms like LOTES and Fositek are working through validation phases, aiming to join the list of qualified suppliers by 2025 to alleviate current shortages in quick disconnect components.
