Shift in Hardware Infrastructure for DeepSeek Training
Reports emerging from online technical communities indicate that DeepSeek is now utilizing Huawei’s Ascend 950 hardware for model training. This development follows a long-standing strategic direction from DeepSeek founder Liang Wenfeng, who previously stated that the company must “step onto the frontier” of hardware development.
The transition to Ascend 950 hardware marks a significant technical pivot for the AI firm. While the move has been discussed extensively on platforms such as Reddit, independent verification of the full scale of this integration remains pending.
Performance Benchmarks and Technical Specifications
Preliminary user reports regarding the performance of the Ascend 950 hardware suggest a focus on high-capacity memory configurations. Discussions on Reddit indicate that 96GB variants of the hardware are being utilized, with some users noting that the memory capacity is competitive for certain workloads.
In unverified performance tests shared by community members, the hardware has demonstrated specific throughput capabilities. One user reported achieving speeds of 30 tokens per second (tok/s) and 60 tok/s using four streams on the Qwen3.8-flash-next model. However, technical inquiries have raised questions regarding the hardware’s performance on dense models, citing potential bandwidth limitations estimated at approximately 200 GB/s.
Users are currently evaluating the Ascend 950 as a potential alternative to other hardware solutions, such as the DGX Spark or Strix Halo, which feature higher bandwidths in the 250–300 GB/s range.
Software Compatibility and Market Context
The integration of Huawei hardware into large-scale AI workflows appears to be driving interest in specialized software forks. Users have noted the necessity of using specific forks of vLLM to optimize performance on these chips. There is also ongoing community testing regarding the compatibility of llama.cpp with the Ascend 950’s 96GB architecture.
The shift toward alternative hardware ecosystems comes amid broader market volatility and fluctuating prices for high-end AI compute resources. For instance, some users have noted significant price increases for DGX Spark units, with reports of price hikes as high as $2,000 within a single week.
As DeepSeek continues to scale its models, the reliance on Huawei’s Ascend series suggests a growing trend of diversifying compute infrastructure away from traditional dominant providers. While the technical community remains divided on the efficiency of these chips compared to established enterprise solutions, the deployment of DeepSeek models on this architecture represents a notable development in the hardware landscape for frontier AI development.