DeepSeek V5 Huawei Ascend chips: reports from Chinese outlets, including STAR Market Daily on November 3, 2026, indicate that DeepSeek plans to train its next V5 model using Huawei Ascend chips, marking an expansion of these processors' use to training rather than only inference.
What happened?
According to supply-chain sources cited by STAR Market Daily and republished by sites such as TechNode and TrendForce, DeepSeek intends to perform V5 pre-training primarily on Ascend 950 chips. Until now, Huawei's Ascend chips have mainly been used to run already-trained models (inference). V5 would be the company's first major model trained natively on this platform.
Neither DeepSeek nor Huawei has officially confirmed the plan. Some reports mention a data center in Ulanqab, Inner Mongolia, and chip volumes in the tens of thousands, but figures vary across sources and remain unconfirmed.
Why does this matter?
Training large models requires far more computing capacity and stability than inference. If the plan materializes, it would be an important test of Ascend chips' maturity for cutting-edge training workloads, amid export restrictions that limit China's access to advanced NVIDIA GPUs.
DeepSeek had already adapted parts of its software stack for Chinese processors and open-sourced components for the Ascend platform. V5 would test whether that adaptation supports a full training run.
What changes in practice?
For China's AI ecosystem, successful V5 training on Ascend would strengthen the narrative of independence from foreign hardware. For the broader market, it reinforces competition among AI chip stacks and pressure on high-end hardware availability.
The reported timeline points to 2027, with possible start of pre-training loads in late 2026, depending on chip availability and software stability.
Image credit: DeepSeek | Source: TechNode / STAR Market Daily
By Grok
