DeepSeek is about to release V4, but this time the most noteworthy aspect is not how large the model's parameters are, but what chip it runs on.
According to The Information, DeepSeek V4 will run on Huawei's Ascend 950PR chip. This is the first frontier AI model built specifically for a domestic Chinese chip architecture.
First, the Model Itself
V4's specifications are impressive:
| Parameter | Value |
|---|---|
| Total parameters | ~1 trillion (MoE architecture) |
| Inference-time activated parameters | ~37 billion |
| Context window | 1 million tokens |
| SWE-bench score | 81% |
The logic behind using the MoE architecture is the same as V3: 1 trillion parameters are stored, but only 37 billion are activated per inference. The actual operational cost is closer to that of a 37-billion-parameter dense model, while the capability can approach the edge of a trillion-parameter model. The post-training API pricing is $0.30 per million tokens, an order of magnitude cheaper than the GPT series.
Multimodality is also integrated: V4 natively supports text, image, and video generation processing.
Why the Huawei Chip Matters
DeepSeek's previous models, including V3 and R1, were trained using NVIDIA A100/H100 (or at least used the CUDA ecosystem). V4 is the first to be built from the ground up on Huawei's CANN architecture.
Ascend 950PR hardware specifications:
- Compute: FP8 1 PFLOPS / FP4 2 PFLOPS
- Interconnect bandwidth: 2 TB/s
- Manufacturing process: SMIC N+3 process (performance close to 5nm class)
- Deployed on Huawei's Atlas 350 accelerator card
To make V4 work on this chip, DeepSeek collaborated with Huawei and Cambricon to rewrite a large amount of underlying code, with the goal of completely bypassing NVIDIA's CUDA ecosystem.
TrendForce's analysis suggests that if this goes smoothly, DeepSeek's development pipeline could achieve substantial independence from CUDA within one to two years.
Alibaba, ByteDance, and Tencent Are All Grabbing Huawei Chips
The market reaction has been direct: Alibaba, ByteDance, and Tencent have placed orders with Huawei totaling hundreds of thousands of Ascend chips. With demand surging, Huawei's chip prices have already risen by about 20%.
Huawei plans to produce approximately 600,000 Ascend 910C units in 2026, double the 2025 figure, bringing total Ascend capacity to 1.6 million units.
From another perspective: if DeepSeek V4 performs well on Huawei chips, domestic AI vendors will have a realistic alternative path — no longer completely dependent on NVIDIA. This has greater implications for the entire industry chain than V4 itself.
But Are Huawei Chips Actually Good Enough?
To be honest, this is still an open question.
The theoretical compute numbers for the Ascend 950PR look decent, but there is no large-scale verified data on its actual distributed training and inference efficiency compared to the NVIDIA H100 ecosystem. V4-Lite is currently in internal testing on API nodes, with developers reporting a 30% improvement in inference speed and noticeable improvements in context recall — but this is in a specially optimized environment.
The real test will be: after the official V4 release, will performance be as stable as advertised?
If it works, this will be a real proof of China's AI hardware independence path. If it doesn't work well, it only shows that the direction of independence is correct, but the timing is not yet right.
V4's official version is expected to be released in mid-April. We'll wait and see the results.
Sources: CocoLoop, DeepSeek's V4 model will run on Huawei chips, The Information reports (WHBL); Decoding DeepSeek V4: How Huawei's Ascend 950PR Is Powering China's Push to Break CUDA Dependence (TrendForce); DeepSeek V4 points to growing use of Huawei chips in AI models (TechWire Asia); DeepSeek V4 And Tencent's New Hunyuan Model To Launch In April (Dataconomy)