[News] DeepSeek V5 Could Complete Training on Huawei Ascend 950 in 2027, Moving Beyond Inference
As Huawei accelerates its Ascend AI chip roadmap, with the Ascend 960DT reportedly slated for early 2027, another milestone may be taking shape. According to Caiwen News, citing industry sources, DeepSeek plans to train its next-generation V5 model primarily on Huawei’s Ascend 950 chips, with the project expected to materialize in 2027.
The move would mark a significant shift for DeepSeek, as Huawei’s Ascend chips have so far been used primarily for inference rather than training, according to Star Market Daily. Sources cited by Caiwen indicated that V5 could become DeepSeek’s first large language model trained natively on Ascend 950 chips.
The growing reliance on Huawei is also evident in DeepSeek’s chip sourcing strategy. According to Caiwen, Ascend remains DeepSeek’s primary training platform, with no reported adoption of solutions from other Chinese chipmakers, including Moore Threads.
Despite these deepening ties, the partnership does not extend to equity investment. Amid market chatters of DeepSeek securing substantial new funding and speculation over a potential IPO, Caiwen’s sources said Huawei holds no stake in the AI startup, with cooperation limited to products, computing resources, and technical optimization.
Why DeepSeek Is Turning to Huawei: A Computing Power Crunch
DeepSeek’s reported shift appears driven more by practical computing constraints than external pressure, according to Caiwen. With U.S. export restrictions limiting access to NVIDIA’s latest high-end AI chips, previously approved H-series GPUs are reportedly struggling to meet the training demands of next-generation models such as DeepSeek V4 and V5, making domestic alternatives increasingly critical.
The potential shift would mark a departure from DeepSeek’s earlier reliance on NVIDIA for model training. As a January Reuters report noted, NVIDIA engineers helped DeepSeek optimize its algorithms, software frameworks, and hardware, with DeepSeek-V3 requiring just 2.788 million H800 GPU-hours for full training.
However, DeepSeek’s transition to Huawei remains in its early stages. According to Caiwen, V5 is still in the selection phase, with its model specifications, key capabilities, and launch timeline undisclosed. Industry sources expect a release sometime in 2027, though the timing will hinge on training progress.
While DeepSeek works toward its V5 training plans, Huawei is also pressing ahead with an accelerated Ascend roadmap. Citing Huawei Rotating Chairman David Wang, Star Market Daily reported that the Ascend 960DT is now expected to be ready in 1Q27, three quarters ahead of its original schedule.
The momentum extends to Huawei’s next-generation chips. According to Guancha, the Ascend 960PR is slated for 3Q27, one quarter ahead of schedule, followed by the Ascend 970 in 2028 and 980 in 2029.
Read more
- [News] Huawei Speeds Up AI Chip Roadmap, Reportedly Pulls Ascend 960DT Forward Three Quarters to 1Q27
- [News] Huawei Says Ascend Market Share Tops NVIDIA in China, Backs NPO Over CPO for Optical Interconnects
(Photo credit: Huawei)
