NVIDIA GB300 Design Updates and Expanded Server Deployments Expected Post-Q3 2025
24 March 2025 · Uncategorized ·
Source: · https://technews.tw/2025/03/18/nvidia-gb300-20253q-expand-shipments/
According to TrendForce’s latest AI server supply chain survey, Nvidia (NVIDIA) is expected to launch the GB300 chip in Q2 of 2025. Compared with the GB200 model, full rack system servers utilizing the GB300 will offer improved computing performance, memory capacity, network connectivity, and power management; consequently, suppliers such as ODMs require additional time for testing and customer validation.
Supply chain dynamics indicate that providers of GB300 components began planning design operations in Q2 this year. Production of both the GB300 chip and Compute Trays is estimated to commence from May onwards, with ODMs currently engaged in early Engineering Sample (ES) stage sample machine designs. TrendForce anticipates full rack system specifications, power requirements, SOCAMM configurations, etc., will be finalized in Q3, enabling mass production and a gradual expansion of GB300 system shipments.
Nvidia’s shipment situation this year remains dominated by the Hopper platform, with Blackwell platforms gradually expanding from Q1 onwards. By Q3, full rack systems are expected to primarily consist of GB200 models. Furthermore, demand for special edition products like H20 has recently increased in China due to factors related to DeepSeek AI.
The upcoming GB300 incorporates several specification updates; notably the NVL72’s network design specifications have been upgraded to meet higher bandwidth requirements and enhance overall computing performance when paired with rack systems. While a battery backup unit (BBU) is not standard for GB300, increasing server cabinet power consumption will likely drive customer adoption of BBU configurations.
Regarding Thermal Design Power (TDP), the mainstream NVIDIA models in 2024 are HGX AI servers featuring TDP ranging between 60kW and 80kW. The currently promoted GB200 NVL72 cabinets, due to significantly increased computational density, have a cabinet-level TDP reaching up to 125kW–130kW. TrendForce predicts that the power consumption of GB300 rack systems will further increase between 135kW and 140kW; most manufacturers are expected to continue utilizing liquid-to-air cooling methods for improved heat dissipation.
In terms of thermal component design, while current GB200 configurations utilize a combined water-cooled plate (cold plate) module incorporating one CPU and two GPUs, the GB300 will adopt individual cold plates per chip. This change increases the value of each cold plate within Compute Tray. Furthermore, due to this shift from integrated modules to independent components, demand for Quick Disconnects (QDs) is expected to increase; currently, QD supply chains are primarily sourced from European and American companies but more Taiwanese manufacturers are anticipated to participate with GB300.
TrendForce notes that expansion of both GB200 and GB300 rack solutions this year will be influenced by several factors. Firstly, lingering effects related to DeepSeek may lead major AI server customers (CSPs) to prioritize cost-effectiveness in their AI investments, potentially driving them towards self-developed ASICs or simpler, more affordable AI server designs. Secondly, uncertainties remain regarding whether the GB200 and GB300 rack supply chains can complete preparations as scheduled; actual progress in supplies and shifts in customer demand will need to be monitored closely.
Supply chain dynamics indicate that providers of GB300 components began planning design operations in Q2 this year. Production of both the GB300 chip and Compute Trays is estimated to commence from May onwards, with ODMs currently engaged in early Engineering Sample (ES) stage sample machine designs. TrendForce anticipates full rack system specifications, power requirements, SOCAMM configurations, etc., will be finalized in Q3, enabling mass production and a gradual expansion of GB300 system shipments.
Nvidia’s shipment situation this year remains dominated by the Hopper platform, with Blackwell platforms gradually expanding from Q1 onwards. By Q3, full rack systems are expected to primarily consist of GB200 models. Furthermore, demand for special edition products like H20 has recently increased in China due to factors related to DeepSeek AI.
The upcoming GB300 incorporates several specification updates; notably the NVL72’s network design specifications have been upgraded to meet higher bandwidth requirements and enhance overall computing performance when paired with rack systems. While a battery backup unit (BBU) is not standard for GB300, increasing server cabinet power consumption will likely drive customer adoption of BBU configurations.
Regarding Thermal Design Power (TDP), the mainstream NVIDIA models in 2024 are HGX AI servers featuring TDP ranging between 60kW and 80kW. The currently promoted GB200 NVL72 cabinets, due to significantly increased computational density, have a cabinet-level TDP reaching up to 125kW–130kW. TrendForce predicts that the power consumption of GB300 rack systems will further increase between 135kW and 140kW; most manufacturers are expected to continue utilizing liquid-to-air cooling methods for improved heat dissipation.
In terms of thermal component design, while current GB200 configurations utilize a combined water-cooled plate (cold plate) module incorporating one CPU and two GPUs, the GB300 will adopt individual cold plates per chip. This change increases the value of each cold plate within Compute Tray. Furthermore, due to this shift from integrated modules to independent components, demand for Quick Disconnects (QDs) is expected to increase; currently, QD supply chains are primarily sourced from European and American companies but more Taiwanese manufacturers are anticipated to participate with GB300.
TrendForce notes that expansion of both GB200 and GB300 rack solutions this year will be influenced by several factors. Firstly, lingering effects related to DeepSeek may lead major AI server customers (CSPs) to prioritize cost-effectiveness in their AI investments, potentially driving them towards self-developed ASICs or simpler, more affordable AI server designs. Secondly, uncertainties remain regarding whether the GB200 and GB300 rack supply chains can complete preparations as scheduled; actual progress in supplies and shifts in customer demand will need to be monitored closely.