TrendForce's latest research, AI server demand drives North America's four major CSPs to accelerate self-developed ASIC chips, and an upgraded version will be launched in one or two years on average. The proportion of China's AI server market will be affected by the export control of new chips in the United States in April. The proportion of chips such as NVIDIA and AMD will drop from about 63% in 2024 to about 42% in 2025. If China's local chip suppliers have China's AI chip policy support, their proportion has increased to 40% this year, almost competing with the proportion of outsourced chips.
TrendForce said that because CSP is gradually increasing its burden on AI work, and plans to reduce the high dependence on NVIDIA and AMD, it is investing heavily in ASIC development in order to achieve cost control, performance and supply chain flexibility and improve operational cost expenditure.
Observing the progress of AI ASICs of the four major American CSPs, Google, which has taken the lead, has launched TPU v6 Trillium, focusing on energy efficiency ratio and optimization of large AI models, and will significantly replace TPU v5 this year. In order to develop new products, Google has added cooperation with Broadcom's one partner model, transforming it into a dual supply chain layout, improving design flexibility and reducing the risk of relying on one supply chain, and also helping to increase the layout of high-level advanced processes.
AWS focuses on Trainium v2, designed in cooperation with Marvell, supports the application of generative AI and large-scale language model training and application, and also cooperates with Alchip to develop Trainium v3. TrendForce estimates that AWS's ASIC shipments will grow significantly this year, with the strongest annual growth performance of the American CSP.
Meta successfully deployed the first self-developed AI accelerator MTIA, and jointly developed the next generation MTIA v2 with Broadcom. Since Meta has a high demand for AI recommendations, MTIA v2 specializes in focusing on energy efficiency optimization and low-latency architecture to ensure both recommended performance and operational efficiency.
The construction of microsoft AI servers is still mainly based on NVIDIA GPU solution, but it has also accelerated the development of ASICs. The Maia series chips mainly optimize the generative AI applications and other services of the Azure cloud platform. The next generation of Maia v2 design has also been decided, and the GUC is responsible for the subsequent physical design and subsequent mass production. In addition to continuing to deepen cooperation with GUC, Microsoft has also introduced Marvell to jointly participate in the design and development of Mai a v2 advanced version, strengthening the layout of self-developed chips, and effectively dispersing development risks.
China's AI supply chain autonomy acceleratesAnalyzing China's independent AI solutions, China is the most important part of the domestic demand market to develop the Shengren chip series, and its application levels include LLM training, local smart city basic construction, and large-scale telecommunications merchant cloud AI applications. National project support and LLM AI applications related to Internet and DeepSeek are booming, and may shake the leading position of China's AI market such as NVIDIA in the long run.
Cambrian's MLU AI chip series is also aimed at applications such as AI training and reasoning for cloud operators. After observing the feasibility of Cambrian's continuous preliminary test and large CSPs in 2024, this year, the Siyuan AI solution will be gradually promoted to the cloud AI market.
TrendForce said that the United States restricts exports of China's high-level AI chips to continue to upgrade, and the Chinese CSP has also accelerated the development of self-developed AI ASICs. Alibaba's T-head launched the Hanguang 800 AI recommendation chip. After Baidu's quantitative production of Kunlun II, it began to develop Kunlun III, focusing on high-performance training and recommendation dual support structures. In addition to its own AI recommendation chip Zixiao, Teng also adopts strategic investment solutions for IC design company Enflame.
Under the background of geopolitics and supply chain emphasis, highlighting the necessity and importance of Chinese chip suppliers, Cambrian and CSPs, investing in self-developed ASICs, and driving the development of two major ecological systems in North Korea and China and Out-of-China.