China is accelerating its homegrown AI hardware push, with plans to sharply expand production of domestic accelerators and reduce reliance on foreign suppliers. According to recent reporting, the country aims to triple output of AI chips within the next few years, positioning local champions such as Huawei, Cambricon, and DeepSeek at the center of the effort.
Beijing’s pivot away from imported AI hardware has been underway for some time, but it is now gathering pace. Huawei is expected to bring a dedicated fab for its AI chips online by the end of the year, with two additional facilities slated to start production next year. Combined, their capacity is projected to be comparable with the output of established foundry players within China, underscoring how quickly domestic manufacturing is scaling.
Meanwhile, leading foundry SMIC is reportedly targeting a doubling of 7nm production by 2026, a sign that demand for homegrown process nodes is surging as AI adoption spreads. The timing matters: a robust pipeline of local manufacturing could help Chinese firms fulfill rapid growth in AI compute needs as export restrictions and global supply constraints persist.
On the product front, China’s latest AI accelerators have closed much of the performance and efficiency gap with global rivals. Huawei’s Ascend 910D and Cambricon’s 690 series are frequently cited as the most competitive domestic options to date. DeepSeek recently said it is tuning its models around the FP8 format to maximize performance and efficiency. While no widely available Chinese accelerator currently advertises native FP8 support, market signals suggest that a local, FP8-capable solution may be imminent.
Policy momentum is adding fuel to the trend. Local authorities in Beijing have reportedly set an objective of achieving complete self-reliance in AI compute, aligning with broader national priorities to reduce exposure to foreign technology. To meet that goal, Chinese firms are pushing forward not only in logic process technology but also in critical complementary areas such as high-bandwidth memory and advanced packaging—pillars needed to build a competitive AI ecosystem.
What to watch next:
– Ramp-up timelines for Huawei’s new fabs and any new tape-outs from Cambricon and other domestic designers
– SMIC’s 7nm capacity growth trajectory and its impact on local accelerator availability
– Announcements of FP8 support or software stacks optimized for lower-precision training and inference
– Progress in domestic HBM and advanced packaging, which are essential for training large AI models
Bottom line: China’s domestic AI chip industry is moving from catch-up to scale-up. Tripling output, expanding local fabs, and advancing memory and packaging technologies could give Chinese AI firms a credible path to compete more directly with established global leaders over the next few years.
