Union consultation: It is expected that the peak shipment of NVIDIA Corporation's (NVDA.US) GB200 cabinet will be delayed to between 2Q25 and 3Q25.

TrendForce Alliance Consulting pointed out in its latest survey that the market is currently focusing on the supply progress of NVIDIA's GB200 rack solution. Due to the GB200 rack's high-speed interconnect interface and thermal design power (TDP) specifications, which are significantly higher than the mainstream market, supply chain operators need more time for continuous adjustment and optimization. It is expected that there will be a chance for mass production no earlier than the second quarter of 2025. NVIDIA's GB rack solution (including GB200, GB300, etc.) is more complex and costly due to the introduction of advanced technology. The main customers are large cloud service providers, Tier-2 data centers, national sovereign clouds, and academic research institutions for HPC/AI applications. With NVIDIA's strong promotion, it is expected that the GB200 NVL72 rack will become the main adoption solution by 2025, with a market share expected to reach near 80%. TrendForce Alliance Consulting stated that in order to enhance the overall computing efficiency of AI/HPC server systems, NVIDIA developed NVLink to provide high-speed interconnection technology between GPU chips. For example, the fifth-generation NVLink used in GB200 significantly exceeds the total bandwidth of the current mainstream PCIe 5.0 technology. In addition, while the HGX AI Server dominated the market in 2024 with a TDP per rack ranging from 60KW to 80KW, the GB200 NVL72 per rack reaches 140KW, doubling the TDP. As a result, operators are trying to expand the use of liquid cooling solutions to address the increased heat dissipation. Due to the higher design specifications of the GB200 rack system, there are market rumors that there may be delayed shipments due to some components not meeting requirements. According to TrendForce Alliance Consulting's survey, the current shipment situation of the Blackwell GPU chip is generally as expected, with only a small amount shipped in the fourth quarter of 2024 and gradually increasing in shipments after the first quarter of 2025. In terms of AI server systems, due to ongoing adjustments in various aspects of the supply chain, the shipment volume by the end of this year is feared to be lower than expected by operators, and therefore, the peak shipment of GB200 rack units in 2025 may be slightly delayed. Traditional air cooling solutions are no longer able to cope with the TDP value of the GB200 NVL72, making liquid cooling technology a necessity. As the GB200 rack solution begins to ship in small quantities at the end of 2024, related operators are also increasing their research and development efforts on liquid cooling components. For example, cooling distribution system (CDU) suppliers are increasing the efficiency of the cooling plates to enhance heat dissipation. TrendForce Alliance Consulting stated that the current cooling capacity of Sidecar CDUs mainly focuses on 60KW to 80KW, with the potential to achieve double or even triple the heat dissipation performance in the future. As for the more efficient liquid-to-liquid (L2L) in-row CDU solution, the heat dissipation capacity has exceeded 1.3MW and is expected to continue to increase to meet the growing demand for computing power.