The Chinese company DeepSeek intends to use Huawei Ascend 950 chips as the main training infrastructure for its next artificial intelligence model, V5. The project has 2027 as its completion horizon, according to information published on Tuesday (3) by Caiwen News, which heard supply chain sources, and reported by TrendForce.
The initiative would represent an important change in the company's infrastructure. Until now, Ascend processors have been used mainly for inference, the stage in which already-trained models process requests and generate responses. With V5, DeepSeek intends to use Huawei's chips also in pre-training, the phase of building the fundamental capabilities of an AI model.
The plan has not yet been officially confirmed by the companies. According to Caiwen News, V5 remains in the stage of selecting its training infrastructure, with no public specifications, defined parameter count or official launch date. Industry sources point to a possible debut in 2027, subject to the progress of development.
DeepSeek and Huawei have already developed tools for training
The possible migration would not start from scratch in terms of software infrastructure. Technical teams from DeepSeek and Huawei have already worked on adapting low-level components during the development of the V4 family, according to the Chinese report.
On September 30, 2026, DeepSeek announced the open-source release of tools for the Ascend platform, including the TileLang compiler, computing libraries and distributed communication components. The collaboration also involved a processing architecture with 128 Ascend 950 chips, as Reuters reported at the time.
These tools make it possible to adapt mathematical operations and coordinate communication between processors, essential tasks for training large models. Native support for Ascend 950 was also incorporated into TileLang, while libraries such as DeepGEMM-Ascend and DeepEP-Ascend began offering components specific to Huawei's hardware.
The availability of these tools, however, does not prove that full V5 training has already begun or that its execution is guaranteed.
DeepSeek also plans to build a computing center in Ulanqab, Inner Mongolia, to house Ascend 950 infrastructure. According to Caiwen News, the facility is still at an early stage, and there is no confirmation that it will host the first stages of V5 training.
The rapprochement comes amid American restrictions on the export of advanced AI processors to China. Sources heard by the Chinese publication say that the capacity of Nvidia chips previously available in the country no longer fully meets the requirements projected for DeepSeek's next models.
If the plan materializes, V5 will be the company's first major model to carry out its pre-training natively with Ascend 950 processors. To date, however, no V5 training results, performance metrics or definitive timeline have been disclosed.
