Jibang Consulting: Blackwell is estimated to account for more than 80% of Nvidia (NVDA.US) high-end GPU shipments in 2025

The Zhitong Finance App learned that according to the latest survey by TrendForce Jibang Consulting, the overall server market has stabilized recently. ODMs all focus on AI server development. Starting in the second quarter, new Blackwell platform products such as NVDA.US (NVDA.US) GB200 Rack and HGX B200 have been gradually released, while the newer B300 and GB300 series have entered the sample delivery verification stage. As a result, Jibang Consulting estimates that Blackwell GPUs will account for more than 80% of NVIDIA's high-end GPU shipments this year.

Looking at recent developments in server ODM, the major North American CSP company Oracle (Oracle) is expanding its AI data center. In addition to mainly bringing order growth to Foxconn (Foxconn), it is also beneficial to businesses such as Supermicro (ultra-micro computer) and Quanta (Quanta). Jibang Consulting expects Supermicro's main growth engine this year to come from AI Server, and has recently won some GB200 Rack projects. Quanta, on the other hand, benefited from the foundation of cooperation with large customers such as Meta, AWS, and Google, and successfully expanded the GB200/GB300 Rack business. In addition, it has obtained an impressive performance in the AI server field recently. Wiwynn (Wiwynn Technology) continues to deepen cooperation with Meta and Microsoft, which is expected to drive performance growth in the second half of this year. It uses ASIC AI Server as its main development force, and is currently the main supplier of AWS Trainium AI models.

AI data center expansion will become the key to scaling up the liquid cooling industry chain

Jibang Consulting said that with the expansion of GB200/GB300 Rack shipments in 2025, the adoption rate of liquid cooling solutions in high-end AI chips continues to rise. Compared with traditional air cooling systems, liquid cooling architectures involve the construction of more complex supporting facilities, such as cabinet and computer room level coolant line layout, cooling water towers, and fluid distribution units (CDU). Therefore, most new data centers currently use the concept of “Liquid Cooling Ready” (Liquid Cooling Ready) at the beginning of design to improve overall thermal management efficiency and expansion flexibility.

Liquid cooling will not only become a standard feature for high-efficiency AI data centers, but will also clearly drive demand for cooling components to heat up, speeding up the pace of supplier shipments. For example, Fositek (Fositek) has officially shipped the GB300 platform's NV QD (quick connector) to match the cold plate (Cold Plate) designed by the parent company AVC for the GB300 NVL72 rack system. The supply chain revealed that FosiTek has mass-produced and shipped quick couplings and floating mounts (floating mounts) required for AWS ASIC liquid cooling, and it is estimated that its quick joint supply ratio on this platform can compete with Danfoss (Danfoss).

Auras (Shuanghong) has also actively deployed the data center liquid cooling market in recent years. Its related business is gradually becoming the core driving force for the company's growth. Its main customers include mainstream server brands such as Oracle, SuperMicro, and HPE (HPE). The product line covers cold water plates and manifold (Manifold) modules. Auras has also begun shipping liquid cooling products to Meta, laying the foundation for entering the GB200 platform liquid cooling system supply chain. It is expected to join Meta and AWS's second-wave cold water plate core supply system in the future.