Connect the Dots: What PCB Designers Should Know About Materials
Elementary, Mr. Watson: The ‘Baton’ Handoff Makes or Breaks the Design
Designers Notebook: Implementing Strategic Planning Principles for PCB Design
TrendForce’s latest findings on the AI server supply chain have revealed that NVIDIA is expected to launch the GB300 chip ahead of schedule in 2Q25. However, due to its improved computing performance, memory capacity, networking, and power management compared to the GB200, ODM partners will require additional time for testing and customer validation.
Recent supply chain developments, GB300–related suppliers are set to initiate design planning in 2Q25. The GB300 chip and Compute Tray are expected to enter production by May, with ODM manufacturers designing early engineering samples. By 3Q25, as rack system configurations, power specifications, and SOCAMM designs are finalized and move into mass production, GB300 full-rack systems are expected to gradually scale up shipments.
Currently, Hopper-based AI servers remain NVIDIA’s primary shipment driver, while the Blackwell platform is gradually ramping up starting in 1Q25. GB200 is expected to be the primary system for full-rack shipments until 3Q25. Meanwhile, demand for the China-specific H20 model has surged, largely due to the DeepSeek effect, which has increased the adoption of AI-optimized hardware in the region.
To support rack-mounted systems, the GB300 NVL72 will feature upgraded networking capabilities to meet higher bandwidth demands and improve overall computational efficiency. Although battery backup units (BBUs) are not standard for GB300, they are expected to see increased adoption as server power consumption continues to rise.
In terms of thermal design power (TDP), NVIDIA’s 2024 mainstream HGX AI servers range between 60 – 80 KW. The GB200 NVL72, currently the flagship system, has a TDP of 125 – 130 KW due to its higher computational density. TrendForce forecasts that the GB300 rack system will see further increases in power consumption—reaching 135 – 140 KW—with most industry players continuing to use Liquid-to-Air cooling solutions to maintain thermal efficiency.
The GB200 cold plate design currently integrates one CPU and two GPUs into a single cooling module. However, with the GB300, each chip will have its own dedicated cold plate, significantly increasing cold plate value in Compute Tray configurations. This shift will also boost demand for quick disconnect (QD) fittings, as each chip will now require separate cooling connections. While GB200 QD suppliers are predominantly Western firms, Taiwanese manufacturers are expected to enter the supply chain for GB300.
TrendForce notes that several factors could influence the deployment of GB200 and GB300 rack systems. The DeepSeek effect is prompting major CSPs to reassess AI investments, potentially favoring custom ASIC development or simpler, cost-effective AI server solutions. Additionally, uncertainties in the GB200 and GB300 supply chain readiness may impact shipment timelines, making actual deployment schedules and end-user demand fluctuations key factors to monitor in the coming months.
