Huawei, Chinese chipmakers adopt DeepSeek V4

Chinese chipmakers including Huawei and Cambricon have begun integrating DeepSeek V4 with domestic hardware platforms in China.

The move comes as companies seek local alternatives amid semiconductor tensions.

Huawei, a Shenzhen-based telecom and chip company, said V4 now runs on its Ascend 950PR and other Ascend processors.

The firm plans to make about 750,000 950PR chips in 2026 after starting mass production in April.

Cambricon, a Beijing-based AI chipmaker, said V4 achieved full-stack support on its platform on release day.

Alibaba, ByteDance, and Tencent have ordered hundreds of thousands of Ascend 950 chips after the model’s launch.

🧠 Food for thought

Implications, context, and why it matters.

DeepSeek V4 focuses on inference efficiency, which keeps answer generation cheaper to run 1.
A hybrid attention setup plus low-precision formats, including a mix of FP8 and FP4, lowers memory use and computing needs during inference 1.
That design, along with support for Huawei’s Ascend platform, makes V4 easier to serve on Huawei NPUs, or neural processing units for AI workloads, as well as Nvidia graphics processing units (GPUs) 1.
DeepSeek says it tested its expert-parallel scheme, which splits model work across several chips, on Nvidia GPUs and Huawei Ascend NPU platforms. That work can require matching numerical precision across both systems 2.
This gives open-weights models, whose trained parameters are publicly available, a believable route beyond Nvidia hardware. It could push Chinese tech firms to use more domestic chips at scale 1.

The U.S. Commerce Department’s Bureau of Industry and Security (BIS), which runs U.S. export controls, says using Huawei’s Ascend 910B, 910C, or 910D chips anywhere in the world may breach those rules 3.
The guidance applies to U.S. and non-U.S. persons worldwide. It warns of criminal, civil, or administrative penalties when required BIS approval was not obtained 4.
These rules could deepen the split in AI infrastructure, with one stack centered on Nvidia’s CUDA, Nvidia’s software platform for AI chips, and another on options such as Huawei’s CANN, Huawei’s AI computing software platform 5.
Domestic substitutes still face speed limits. CFR, the Council on Foreign Relations, is a U.S. think tank. It estimates that by 2027 the best U.S. AI chips could be about 17 times more powerful than Huawei’s best products 6.