Nomura stated that the risk of slow shipment growth for NVIDIA's GB200 AI server chips has been largely digested by the market, and the supply chain is accelerating its transition to the next-generation GB300 architecture. Nomura analyst Anne Lee noted in a report on February 12 that the yield of GB200 racks has improved after the Spring Festival, but overall shipments remain below expectations. The market has largely digested the risk of slow GB200 rack shipments in the first half of 2025: Some sell-side analysts have recently lowered their expectations for GB rack shipments in 2025 from the previous 35,000-50,000 units to 20,000-25,000 units. Nomura currently expects global server revenue to grow by 46% year-on-year in 2025 and by 22% in 2026. Among these, AI server revenue is expected to grow by 75% and 31% in 2025 and 2026, respectively. These forecasts are generally consistent with the recent capital expenditure guidance raised by major U.S. cloud service providers (CSPs). Accelerating GB300 Progress, BOM Design Still Faces Challenges Nomura expects that NVIDIA's chip and module-level supply chain will begin to accelerate its transition to B300 and related modules starting in the second quarter of 2025. They estimate that by the second quarter of 2025, over 50% of NVIDIA's chips and modules will be based on B300. The report believes that there has been no significant change in the downstream progress of B300 and GB300. NVIDIA's supply chain may begin mass production of B300 OAM (or SXM) modules starting in the second quarter of 2025. The B300 HGX is expected to launch in mid-2025, while the GB300 system (standard version) may see small-scale market entry in the third quarter of 2025. However, B300 and GB300 still face challenges. Nomura Securities believes that the bill of materials (BOM) design for B300 UBB and GB300 computing boards has not yet been finalized, as NVIDIA is trying to introduce a large number of new component suppliers to reduce costs and achieve supply diversification. Huge Supply-Demand Gap for CoWoS The report pointed out that there are differences between the various levels of NVIDIA's AI server supply chain (from chips to modules to final racks). In terms of upstream CoWoS chip supply, Nomura Securities has adjusted the supply of Hopper and expects that all CoWoS-L will be used for the production of B200/300 in 2025. In terms of demand, Nomura Securities estimates that only 4.4 million and 5.5 million NVIDIA GPUs will be made into modules and systems in 2024 and 2025, respectively. There is also a significant gap between the ideal number of GB racks and the final GB rack shipment volume. Nomura Securities believes that the supply-demand gap for NVIDIA GPUs from 2024 to 2026 is about 20-23%. Optimistic About ASIC AI, Focus on AWS Product Transition Nomura Securities reiterated its optimism about the ASIC (Application-Specific Integrated Circuit) AI field and believes that Amazon (AWS) will actively expand the scale of its own ASIC AI servers in 2025. They expect the supply of AWS's Trainium 2 chips to peak in the second quarter of 2025 and accelerate its decline in the fourth quarter The report points out that AWS is eager to migrate to Trainium 3 as soon as possible (targeting the fourth quarter of 2025). Nomura Securities believes that the market may focus on the beneficiaries of Google's TPU, as well as those companies that benefit from AWS Trainium 3