DeepSeek V5 Plans to Train on Ascend 950 in 2027, Marking Domestic Chips’ Entry into Frontier Model Training

DeepSeek plans to use Huawei’s Ascend 950 chips as the core training hardware for its V5 model in 2027.

Fact Summary

According to a November 4, 2026 report by Digitimes, DeepSeek is planning to expand its cooperation with Huawei from inference to full model training, with Ascend 950 chips set to become the main hardware for V5 model training. The report notes that the arrangement may begin in 2027 and involves adjusting clusters from inference use to training use.

Mechanism Breakdown

The report indicates that DeepSeek has previously used Huawei chips in inference scenarios; the shift to training means the chips will need to handle larger-scale computational workloads. Ascend 950 is positioned as an Nvidia alternative, and its performance in training will be a key test of whether domestic chips can support frontier models. A deployment plan related to a data center in Ulanqab, Inner Mongolia, is mentioned, but specific scale details are not fully elaborated in public reports.

From a computational workload perspective, inference tasks usually focus on low-latency responses for single forward passes, while training involves large-scale gradient computation and iterative parameter updates. The two differ fundamentally in computing resource scheduling, memory bandwidth, and cluster interconnection requirements. This shift requires Ascend 950, building on its existing inference optimization, to further handle the pressure of distributed parallel training; the cluster adjustment itself is a direct test of hardware stability and scalability.

Although the data-center-level deployment plan does not disclose specific scale, it hints at the linkage between physical infrastructure and chip performance. Training tasks depend far more than inference scenarios on power supply, cooling efficiency, and network topology. The Ulanqab site may have already taken these long-term constraints into account, reflecting an overall consideration from hardware selection to system-level adaptation.

Industry Impact

If training tasks proceed smoothly, Huawei’s Ascend series will move from the edge inference field into mainstream model training competition. In the global AI chip market, Nvidia’s dominant position may face more pressure from localized alternatives. The trend toward diversified computing power choices among Chinese AI teams will become more evident, and companies upstream and downstream in the supply chain will need to reassess hardware adaptation and software ecosystem support.

This shift is directly related to the hardware foundation for subsequent model iterations.

At the supply chain level, hardware adaptation is no longer limited to lightweight integration of inference frameworks; training frameworks will need deep optimization for the Ascend 950 instruction set and communication protocols. Software ecosystem support will thus become a bottleneck, and the maturity of developer toolchains, debugging environments, and performance tuning methods will determine the actual pace of implementation. Once diversified choices form path dependence, the risk of upstream and downstream companies relying on a single supplier will also be redistributed.

Strategic Assessment

[Analysis] If Ascend 950 proves feasible in V5 training, it will provide a reference case for other domestic chips entering the training market, but the actual effect still needs time to be tested. Current reports have not disclosed specific performance metrics or a completion timetable.

The formation of a reference case requires multidimensional validation: not only single-chip peak computing power, but also multi-card interconnection efficiency, fault recovery mechanisms, and long-term stability. If other domestic chips use this as a benchmark, they can plan ahead in architecture design, compiler optimization, and other areas to reduce trial-and-error costs.

However, the lack of public performance metrics and timelines means the strategic assessment remains in the observation stage. Model training cycles are long and iteration risks are high; any uncertainty at the hardware level could affect the overall R&D rhythm. If DeepSeek and Huawei’s cooperation proceeds as planned, it will provide the industry with a replicable hardware-software collaboration paradigm; if deviations occur, it could slow the overall penetration of domestic chips in the training field.