On November 3, 2026, TrendForce relayed a Caiwen News report based on unnamed industry sources: DeepSeek was considering Huawei Ascend 950 as the core training compute for V5. Caiwen said the model was still in chip selection, making Ascend 950 a reported plan rather than a settled hardware choice.
What the report says about V5 and Ascend 950
The sources cited by Caiwen described Ascend 950 as the planned core platform for training DeepSeek V5 and said the model could become DeepSeek’s first large language model trained natively on Ascend. The report concerns model training, not simply running an existing model to generate responses.
Caiwen’s sources also forecast a possible V5 release in 2027, with timing dependent on training progress. That is a broad, conditional expectation, not a release date.
The reported engineering work behind the plan
Caiwen reported that DeepSeek’s V4-stage work with Huawei included adapting low-level operators and customized development for Ascend. Operators are software routines used in a model’s computations. Caiwen’s sources also described moving operator paths originally developed with CUDA into an open-source Python-operator ecosystem.
Those details describe reported preparation for using Ascend. They do not describe a V5 training result. The reported hardware plan remains at the chip-selection stage.
Why training would be a different use of Ascend
Inference runs a trained model to produce outputs; training uses compute to build or update the model. Caiwen’s sources characterized DeepSeek’s earlier use of Ascend as inference and described V5 as a possible shift to training. If that plan proceeds, it would mark a different role for the platform in DeepSeek’s model development.
Caiwen also described DeepSeek’s planned computing center in Ulanqab as being in an early construction phase. The report did not identify it as V5’s initial training site.
