Three of China's largest tech companies are vying for the same chip. Alibaba, ByteDance, and Tencent have placed bulk orders with Huawei for the Ascend 950PR, totaling hundreds of thousands of units. Mass production of these chips only began this month, and Huawei's annual shipment target is approximately 750,000 units. This concentrated procurement by the three giants has driven the price of the 950PR up by 20% in recent weeks.
The trigger for this buying frenzy was DeepSeek V4. The three companies planned to distribute the models to businesses and developers through their cloud services after the official release of V4, and integrate them into their respective AI applications.
The 950PR is priced at around 50,000 RMB (about 6,900 USD), while the high-end version equipped with HBM memory rises to 70,000 RMB, but even so, it still can't stop this wave of purchases.
DeepSeek V4 rejects Nvidia and prioritizes Huawei compatibility.
Behind this wave of orders lies a larger strategic signal. According to an exclusive report by Reuters on February 25, DeepSeek only opened an early access window to Chinese chip companies such as Huawei before the release of V4, explicitly rejecting the participation of NVIDIA and AMD.
The usual practice is for chip companies to obtain the large-scale model in advance before its official release, in order to prepare supporting software and optimization tools. DeepSeek's choice this time gave Huawei a software adaptation advantage before the public release of V4, while Nvidia was completely excluded.
DeepSeek has also been collaborating with Huawei and chip design company Cambricon to advance hardware optimization for V4.
DeepSeek V4 Specifications Highlights
DeepSeek V4 employs a MoE (Mixture-of-Experts) architecture, with a total of approximately 1 trillion references, but only uses about 37 billion references per inference, effectively maintaining low latency and low cost. The model supports multimodal input including text, images, and code, with a context window of up to 1 million tokens, and achieves a score exceeding 80% in the SWE-bench code benchmark.
According to NxCode 's estimates, the V4 API is priced at approximately $0.14 per million input tokens, which is 20 to 50 times cheaper than leading Western models.
V4 was originally scheduled for release in February 2026, but has been repeatedly delayed due to the need to rewrite the underlying code when migrating from the NVIDIA architecture to Huawei chips. DeepSeek is currently developing two additional V4 variants, each optimized for different capabilities and both designed for Chinese chip architectures.
V4 is expected to be released within weeks.
