China Aligns Domestic AI Models with Custom Chips
China's technology ecosystem is rapidly integrating its domestic large language models, such as DeepSeek and Z.AI, with homegrown semiconductor designs. This strategic alignment is seen as a direct policy and technological response to U.S. export controls. The effort aims to create a self-sufficient AI technology stack independent of Western hardware and software.
- U.S. export controls, which began escalating in October 2022, block China from accessing top-tier AI chips like Nvidia's A100 and H100, as well as the advanced equipment needed to manufacture them. This has directly spurred China's push for a self-reliant hardware and software ecosystem. - In response, Chinese tech firms have formed industry groups like the "Model-Chip Ecosystem Innovation Alliance," which includes AI model developers like StepFun and chip manufacturers such as Huawei, Biren, and Moore Threads, to standardize and unify the domestic technology stack. - Huawei's Ascend 910B is a key domestic alternative to Nvidia's restricted chips and is now used to train nearly 50% of China's large language models. While still generations behind Nvidia's most advanced products, the 910B is reportedly competitive with the Nvidia A100, a chip widely used before the latest U.S. restrictions. - The performance gap is significant at the high end; Nvidia's H100 boasts over five times the FP16 computing power of Huawei's Ascend 910B. Furthermore, Chinese chips lag in cluster expansion capabilities, with domestic interconnect bandwidth being less than a quarter of Nvidia's NVLink, leading to lower efficiency in large-scale training. - To overcome manufacturing limitations imposed by sanctions, Chinese researchers are exploring alternative architectures. One such avenue is photonic AI chips, like "LightGen" and "ACCEL," which use light instead of electrons for computation and claim up to 100 times greater speed and efficiency for specific AI tasks compared to an Nvidia A100. - DeepSeek's open-source models are designed to be less reliant on high-end hardware, making them compatible with less powerful domestic chips. This software innovation allows for comparable AI performance with more cost-effective hardware, driving adoption among Chinese tech firms, including major telecom operators and automakers. - The Chinese government is heavily subsidizing this push through initiatives like the "Made in China 2025" plan and its 15th five-year plan (2026-2030), which prioritize semiconductor and AI self-sufficiency. This includes adding domestic AI chips from companies like Huawei and Cambricon to official government procurement lists for the first time. - This strategy extends beyond hardware to a complete, vertically integrated supply chain, with Huawei and its partners developing everything from chip design software (EDA) to advanced packaging techniques and the production of essential components like high-bandwidth memory.
