DeepSeek V4 Runs on Huawei Ascend 950 Chips

Beijing's flagship AI model is now adapted to Huawei's Ascend silicon, marking a milestone in China's self-reliant AI stack.

DeepSeek, one of China's leading open-source AI model developers, announced that DeepSeek V4 has been successfully optimized to run on Huawei's Ascend 950 series AI accelerators. This development represents a critical step toward building a domestically self-sufficient artificial intelligence ecosystem in China.

A Strategic Partnership

The collaboration between DeepSeek and Huawei marks a significant convergence of China's two most important AI infrastructure components: cutting-edge foundation models and locally manufactured AI compute hardware. DeepSeek V4, which features 671 billion parameters with a mixture-of-experts architecture, has been specifically adapted to leverage the Ascend 950's architecture.

According to sources familiar with the project, the optimization work took approximately six months and involved close coordination between DeepSeek's model training team and Huawei's Ascend compute platform engineers. The result is a version of DeepSeek V4 that can run inference workloads with performance comparable to Nvidia H100-powered deployments, but entirely on Chinese silicon.

Technical Milestones

The Ascend 950, Huawei's latest AI accelerator chip, features a CANN (Compute Architecture for Neural Networks) framework that required significant software adaptation work. DeepSeek's engineering team rewrote critical portions of their model's attention mechanisms and tensor operations to align with CANN's compute paradigms.

Benchmark tests conducted in late March 2026 showed that DeepSeek V4 running on Ascend 950 clusters achieved:

• 82% throughput efficiency compared to baseline H100 configurations
• Inference latency within 15% of international benchmarks
• Cost per million tokens reduced by approximately 40% due to domestic pricing

Implications for China's AI Strategy

This development arrives at a critical juncture for China's artificial intelligence ambitions. With ongoing export restrictions limiting access to advanced Western AI chips, the successful pairing of DeepSeek V4 with Ascend hardware demonstrates that China is making meaningful progress toward technological self-reliance in AI infrastructure.

Industry analysts note that the collaboration also sends a signal to other Chinese AI model developers that Huawei's Ascend platform is maturing into a viable alternative for production-grade AI workloads. Several other Chinese foundation model companies are reportedly in discussions with Huawei to explore similar optimization partnerships.

Market Impact

The announcement has already generated significant interest among Chinese enterprises evaluating AI deployment strategies. Companies that were previously hesitant to commit to Ascend-based infrastructure due to limited model compatibility now see a clearer path forward.

DeepSeek has indicated that it will offer DeepSeek V4 as an Ascend-optimized API service through its commercial platform starting in Q2 2026. Pricing is expected to be competitive with existing GPU-based offerings, with the added benefit of guaranteed domestic availability regardless of geopolitical factors.

Looking Ahead

While this milestone is significant, challenges remain. The broader AI ecosystem still relies heavily on software frameworks, developer tools, and training pipelines that were originally designed for CUDA-based architectures. Huawei and its partners are actively working to bridge these gaps, but full ecosystem maturity will likely require another 12-24 months of concentrated effort.

Nevertheless, the successful deployment of DeepSeek V4 on Ascend 950 represents tangible proof that China's domestic AI stack is advancing beyond proof-of-concept demonstrations toward real-world commercial viability.