China to triple AI chip production by 2026 for DeepSeek models
China aims to triple domestic AI chip output in 2026 as part of a government-backed plan to supply compute power for leading models like DeepSeek while reducing dependence on Nvidia under U.S. export restrictions.
The country aims to increase AI accelerator output roughly threefold next year by adding new manufacturing capacity focused on Huawei and expanding foundry access, Financial Times reports.
Semiconductor Manufacturing International Corp. (SMIC) plans to double its 7nm production capacity, providing wider access for domestic GPU designers, including Cambricon, MetaX and Biren. Memory maker ChangXin Memory Technologies (CXMT) is preparing high-bandwidth HBM3 components to work with these processors.
DeepSeek is adapting its software to work with Chinese hardware. The startup's V3.1 release added an FP8 precision format that the company says is "optimized for soon-to-be-released next-generation domestic chips." The format reduces memory usage and speeds up inference on local accelerators. The firm also introduced a hybrid inference mode that allows users to activate deeper reasoning when needed.
However, the transition faces challenges. In August 2025, DeepSeek delayed launching a new model after training attempts on Huawei chips failed to meet performance standards. The delay highlighted the performance gap between Chinese chips and Nvidia components.
Chinese chip companies are benefiting from increased demand. Cambricon reported a swing to record first-half profit as Chinese platforms test domestic accelerators and software stacks compatible with models like DeepSeek. The results reflect rising demand amid restricted access to U.S. technology.
Beijing is providing policy support through targeted funding for research, subsidized compute resources, and closer coordination between AI model developers and hardware suppliers. The goal is to accelerate self-reliance while closing performance gaps with U.S. technology.
The 2026 capacity expansion would give Chinese AI developers more consistent access to training and inference chips designed for FP8-heavy workloads that DeepSeek is promoting, even as U.S. export restrictions continue limiting advanced Nvidia shipments.
Author
Has been exploring the enigmas of the crypto industry since 2017, transforming them into accessible narratives. Relies on dark chocolate and nuts as a secret source of energy and inspiration.
