According to the latest AI Server industry research from TrendForce, the ongoing construction of AI infrastructure is driving continuous innovation in AI chips from companies like NVIDIA, AMD, and Google. The Thermal Design Power (TDP) of single chips has generally surpassed 1kW, and the power of full-rack AI Server solutions has climbed to hundreds of kW. As a result, liquid cooling is gradually becoming the standard configuration for high-end AI infrastructure. TrendForce estimates that the penetration rate of liquid cooling in AI chips will reach 53% by 2026, approaching 60% by 2027.
High-end chips from NVIDIA, AMD, and Google are driving the demand for liquid cooling. TrendForce indicates that by 2026, accelerated AI investments from Tier 2 data center operators and increased demand from Chinese cloud service providers (CSPs) for overseas AI projects are expected to double shipments of full-rack solutions like NVIDIA GB/VR. Although the Kyber NVL144 timeline may be delayed in 2027, strong investment in AI infrastructure means overall GPU rack demand remains unaffected, with shipments projected to grow over 30% year-on-year.
AMD's overall strategy will expand from single GPU products to a comprehensive AI platform layout. Starting in the second half of 2026, the company will focus on the Helios AI rack solution, which combines CPU, GPU, and high-speed interconnects, with mass production expected in 2027. Additionally, AMD is advancing the MI450 platform, with plans to introduce the next-generation MI500 platform, positioning it to compete with NVIDIA's Rubin Ultra. TrendForce points out that these chip solutions have fully adopted liquid cooling technology. As AI chips evolve and CSPs accelerate the upgrade of AI data center architectures, the penetration rate of liquid cooling in AI chips is expected to rise from approximately 33% in 2025 to 53% in 2026.
Among major CSPs, Google is the most active in adopting liquid cooling, with coverage exceeding 80% of its AI Servers. Leveraging years of experience in developing its own AI accelerator (TPU) and integrating AI infrastructure, Google has established a highly customized liquid cooling architecture. In response to rising TPU power consumption and rapid shipment growth, Google has been the first among CSPs to extend liquid cooling from single server-level designs to full-rack system designs, becoming another key driver for the overall expansion of liquid cooling penetration.
From a supply chain perspective, liquid cooling systems include key components such as cold plates, manifolds, and cooling distribution units (CDUs). Compared to air cooling systems, these significantly improve heat dissipation efficiency while enhancing Power Usage Effectiveness (PUE). With the full adoption of the fanless, all-liquid cooling architecture in NVIDIA's Vera Rubin platform, liquid cooling configurations have expanded from primarily covering GPU/CPU to include CX9 network cards, busbar/power boards, optical transceiver modules, and other critical components, thereby increasing the per-rack cooling value. Current major cold plate suppliers include Cooler Master, AVC, BOYD, and Auras. Among them, AVC is expected to begin shipping cold plate modules for Vera Rubin in the third quarter and has already entered the AI ASIC platform supply chains of AWS, Google, Meta, and OpenAI. Regarding vapor chambers co-packaged with chips, NVIDIA originally planned to use a two-piece design for the Rubin platform to improve thermal efficiency. However, due to issues like substrate warpage, it has temporarily transitioned to a one-piece solution. The vapor chambers for the Rubin platform are currently exclusively supplied by Jentech.
Comments