Chinese AI company DeepSeek is reportedly preparing to train its upcoming V5 language model entirely on Huawei hardware, potentially reducing its dependence on Nvidia's computing infrastructure.
The model, expected to contain approximately eight trillion parameters, could become a major test of China's ability to develop advanced AI systems using domestically manufactured processors and software.
According to industry reports, DeepSeek has ordered approximately 160,000 Huawei Ascend 950DT chips for a planned 1GW data center in Inner Mongolia.
Deliveries are expected to increase during the fourth quarter of 2026, although neither company has officially confirmed the complete training arrangement.
Huawei Ascend 950DT Could Power DeepSeek's Largest AI Training Project
Huawei's Ascend 950DT accelerator uses the company's third generation DaVinci architecture and is reportedly manufactured using SMIC's N+3 process, an advanced 7nm class technology.
The processor supports AI training and inference workloads, including lower precision computing formats such as FP8, FP4 and Huawei's HiF8.
Each chip reportedly costs approximately 111,000 yuan, equivalent to around $16,000.
Based on that estimate, DeepSeek's proposed order could reach $2.56 billion.
| Specification | Huawei Ascend 950DT |
|---|---|
| Architecture | DaVinci v3 |
| Manufacturing process | SMIC N+3 |
| Memory | 144GB HiZQ 2.0 HBM |
| Memory bandwidth | 4TB/s |
| Interconnect bandwidth | 2TB/s |
| Estimated chip price | $16,000 |
| Reported DeepSeek order | 160,000 chips |
Huawei's Atlas 950 SuperPoD infrastructure can connect up to 8,192 accelerators, providing a reported peak computing capacity of 8 exaflops using FP8 calculations.
DeepSeek Is Developing Software Alternatives to Nvidia CUDA
Hardware availability represents only part of DeepSeek's strategy.
The company has reportedly adapted several AI optimization tools for Huawei's Ascend architecture, including TileLang, DeepGEMM, DeepEP and FlashMLA.
TileLang is particularly important because it allows developers to write optimized computing code without depending directly on Nvidia's CUDA ecosystem.
Recent optimization work reportedly improved model computing utilization to approximately 45%, compared with around 30% previously reported for Huawei's Ascend 910C.
However, performance under isolated testing conditions does not guarantee equivalent efficiency during complete model training.
DeepSeek V5 Could Test China's Domestic AI Infrastructure
Successfully training V5 would demonstrate whether Huawei's hardware and DeepSeek's software tools can support a complete large language model training process.
It could also encourage more Chinese AI developers to consider domestic accelerators.
However, the reported plan remains unconfirmed, and its success will depend on hardware reliability, software optimization and sustained performance across thousands of processors.
For now, DeepSeek's proposed V5 training project represents an important potential step toward reducing China's dependence on Nvidia for advanced AI development.
Discussion (0)
Be the first to comment.