NVIDIA is set to unveil its new line of AI accelerators, the B300 and GB300 AI, at the 2025 GTC conference. According to Taiwan Economic Daily, the next-generation architecture, dubbed 'Blackwell Ultra,' promises to offer significant improvements, including a boost in performance of up to 1.4 times thanks to an increased power consumption of 1400W TDP, a 40% increase over the current GB200 model's 1000W. The computational power is expected to increase, with an approximate 1.4-fold rise in FP4 performance. Additionally, the memory capacity will increase, from 192GB to 288GB, thanks to 12-layer HBM3E stack modules. This will enhance the capabilities of servers when executing complex AI tasks. The supply chain will be led by Foxconn and Quanta, which have been actively researching and developing the new architecture. A significant enhancement is that the GB300 servers will transition to liquid cooling, expected to boost the system's efficiency. Although the official release date for the B300 series has not been announced, a possible launch date is forecast for the end of 2025.