US and Chinese CSPs Accelerate In-House AI Chip Development

TrendForce reports that geopolitical tensions are driving US and Chinese cloud service providers (CSPs) to accelerate the development of in-house AI chips. This shift aims to reduce reliance on imported chips from companies like NVIDIA and AMD. In the US, major CSPs are releasing new iterations of application-specific integrated circuits (ASICs) every one to two years to manage growing AI workloads and optimize operational costs.

In China, the AI server market is adapting to new US export controls, which are expected to decrease the share of imported chips from 63% in 2024 to around 42% in 2025. Domestic chipmakers such as Huawei are projected to increase their market share to 40%, supported by government policies promoting homegrown AI processors.

Among US CSPs, Google leads with its TPU v6 Trillium, enhancing energy efficiency and performance for large-scale AI models. AWS focuses on its Trainium series, co-developed with Marvell, while Meta is advancing its MTIA series with Broadcom. Microsoft is also progressing with its Maia series, collaborating with Marvell to strengthen its chip design capabilities.

In China, Huawei is developing its Ascend series of AI chips, targeting domestic needs such as LLM training and AI-powered telecom networks. Cambricon is expanding its Siyuan chip series, and major CSPs like Alibaba, Baidu, and Tencent are advancing their ASIC initiatives to enhance AI training and inference capabilities.