The three Chinese tech giants are competing for the same chip. Alibaba, ByteDance, and Tencent have placed bulk orders for Huawei's Ascend 950PR, totaling hundreds of thousands of units. The chip just entered mass production this month, with Huawei targeting annual output of about 750,000 units. The scramble has driven the 950PR's price up by 20% in recent weeks.
The rush is triggered by DeepSeek V4. All three plan to distribute the model through their cloud services to enterprises and developers after the official launch, integrating it into their AI applications. The standard 950PR is priced at around 50,000 yuan (about $6,900), with the HBM-equipped version going for 70,000 yuan, yet demand shows no sign of cooling.
DeepSeek V4 Shuns NVIDIA, Gives Huawei a Head Start
Behind the orders lies a bigger strategic move. According to a Reuters exclusive on February 25, DeepSeek opened its pre-release window solely to Chinese chipmakers like Huawei, explicitly excluding NVIDIA and AMD. Normally, chipmakers get early model access to prepare software and optimization tools. DeepSeek's decision hands Huawei a lead in software adaptation while locking out NVIDIA entirely. DeepSeek has also been collaborating with Huawei and Cambricon on V4 hardware optimization.
DeepSeek V4 Spec Highlights
DeepSeek V4 uses a Mixture-of-Experts (MoE) architecture with about 1 trillion total parameters, but activates only 37 billion per inference, keeping latency and cost low. It supports multimodal input — text, images, and code — with a context window of 1 million tokens. Its SWE-bench coding benchmark score exceeds 80%. According to NxCode, the API pricing is estimated at $0.14 per million input tokens, 20 to 50 times cheaper than Western frontier models.
Originally slated for February 2026, V4 has been delayed multiple times due to the need to rewrite underlying code while migrating from NVIDIA's architecture to Huawei chips. DeepSeek is also developing two additional V4 variants, each optimized for different capability dimensions, all designed for Chinese chip architectures. V4 is expected to launch within weeks.