ByteDance, Alibaba rush for Huawei AI chips after DeepSeek

Chinese internet firms including ByteDance, Tencent, and Alibaba are seeking more Huawei Ascend 950 AI chips after Chinese AI startup DeepSeek released its V4 model, which runs on the chips.

Cloud providers and GPU rental firms are also placing orders as US export controls limit Chinese access to Nvidia’s processors and the H200 has yet to reach China.

Huawei expects to ship about 750,000 Ascend 950PR chips in 2026.

Supply may still fall short because restrictions on advanced chipmaking tools limit domestic production.

🔗 Source: Reuters

🧠 Food for thought

Implications, context, and why it matters.

DeepSeek’s V4 preview suggests domestic chips can support advanced AI at lower cost

Demand for Huawei chips rose after DeepSeek previewed its V4 model and said V4-Pro can rival closed-source models from OpenAI and Google 1.
DeepSeek priced V4-Pro output at US$3.48 per million tokens, compared with US$30 from OpenAI for the same volume 1.
That mix of claimed performance and lower pricing, plus Huawei’s pledge of “full support” for DeepSeek models on Ascend processors, strengthens the case that Ascend could replace some Nvidia demand as U.S. export controls restrict Chinese access to the most advanced chips 1.
DeepSeek released the V4 series under the MIT License, a permissive open-source software license that allows broad reuse, which gives developers wider access to the models 2.

A split AI market could challenge U.S. tech dominance

A strong model running on domestic hardware suggests U.S. export controls pushed Chinese AI developers to work with tighter computing limits and build more efficient systems 1.
DeepSeek has expanded in Africa with cheaper, less power-hungry AI offerings alongside Huawei’s storage and cloud computing services 2.
Nvidia CEO Jensen Huang warned that DeepSeek launching first on Huawei would be “a horrible outcome” for the U.S. 1.
DeepSeek’s rise also faces allegations from U.S. officials, OpenAI, and Anthropic over large-scale copying of American AI models 1. The claims also cover improper “distillation,” a technique for training one model using another model’s outputs 1.