🧠 Food for thought

Implications, context, and why it matters.

Sanction-era chips rely on local factories and state money

• Huawei’s Ascend 910C chip uses a compute chiplet (a smaller processor tile used as a building block in a larger chip) made by China’s SMIC (Semiconductor Manufacturing International Corp., China’s largest contract chipmaker) on its 7nm-class N+2 process, a domestic step forward under U.S. sanctions and limited access to leading-edge TSMC (Taiwan Semiconductor Manufacturing Co., the world’s biggest contract chipmaker) process technologies 1.
• Manufacturing output has improved. Production yields (the share of usable chips from a wafer) reportedly doubled over the past year to nearly 40%, with plans to make more than 100,000 units by the end of 2025 2.
• Policy also props up demand. Shenzhen has a 4.5 billion yuan ($630 million) incentive plan that gives businesses vouchers covering up to 60% of computing power costs, capped at 10 million yuan 3.

China’s AI hardware progress centers on inference

• The Ascend 910C competes best in AI inference (running already trained models). DeepSeek (a Chinese AI research lab) researchers estimate it reaches about 60% of Nvidia H100 inference performance 1.
• AI training (building and refining models) still favors Nvidia. Tom’s Hardware (a technology news and review site) summarized DeepSeek’s findings and said the Ascend 910C is not the top pick for training 1.
• Chinese processors still struggle with long training reliability, a gap tied to Nvidia’s tightly linked hardware and software ecosystem built over two decades 1.
• Huawei is also building its in-house CANN (Compute Architecture for Neural Networks) software ecosystem as an alternative to Nvidia’s CUDA (Compute Unified Device Architecture, Nvidia’s dominant software platform for GPU computing) 2.