Cambricon targets triple chip output to rival Nvidia in China

Cambricon Technologies plans to more than triple its AI chip production in 2026, aiming to fill the gap left by Nvidia’s exit from the Chinese market and compete with Huawei.

The Beijing-based chip designer is targeting delivery of about 500,000 AI accelerators next year, including up to 300,000 units of its advanced Siyuan 590 and 690 models, according to people familiar with the matter.

Cambricon will use Semiconductor Manufacturing International Corp.’s (SMIC) “N+2” 7-nanometer process, though current yields for these chips are about 20%, meaning most chips produced are unusable.

Goldman Sachs estimates Cambricon will build 142,000 AI chips in 2025.

🔗 Source: Bloomberg

🧠 Food for thought

Implications, context, and why it matters.

SMIC’s 7nm yields will determine if Cambricon can meet 2026 targets

Cambricon hits 500,000 units only if Semiconductor Manufacturing International Corp. (SMIC) lifts N+2 (an enhanced 7nm variant) yields from about 20% to near 60–70%, a level analysts say SMIC reached on standard 7nm after long tuning 1.
SMIC runs about 20,000 7nm wafers each month, while Huawei takes roughly 15,000, so little headroom remains for Cambricon without more tools and space 1.
SMIC plans to double 7nm output to chase AI chips 2, yet Huawei plus Cambricon’s 300,000 advanced parts may still overwhelm supply if yields stay low.
N+2 relies on multiple-patterning Deep Ultraviolet (DUV) gear because SMIC lacks Extreme Ultraviolet (EUV) access 3, which raises cost and slows scaling.

Software tooling gaps give PyTorch vendors an opening

Cambricon’s torch_mlu plugin for its Machine Learning Unit (MLU) accelerators often needs workflow changes like custom activations and quantization APIs, which deters PyTorch users used to Nvidia’s CUDA flow 4.
System integrators with MLOps (machine learning operations) teams can help firms like ByteDance, Alibaba move PyTorch models onto MLU cards when operator coverage falls short 4.
Compiler specialists can win work because Cambricon still leans on manual edits plus fusion passes 4, while PyTorch uses CUDA with the CUDA Deep Neural Network library (cuDNN) to cover these steps 5.
Enterprise software vendors can ship tools that automate quantization and deployment for Cambricon hardware, cutting the heavy setup seen in its porting guides 4.