DeepSeek plans to train its V5 model on Huawei Ascend 950 chips

DeepSeek plans to train its V5 model on Huawei’s Ascend 950 chips next year. That is a big step for a Chinese AI lab that has reportedly leaned on Nvidia hardware for its most demanding work.

What DeepSeek is reportedly building

The plan centers on a new data center in Ulanqab, Inner Mongolia. DeepSeek reportedly intends to deploy at least 160,000 Huawei Ascend 950DT chips at the site.

That fleet was initially meant for inference, according to the reports. Inference is the work of answering user prompts with a model that has already been built.

The more ambitious shift is toward pre-training. That is the long, compute-hungry stage where a model learns from enormous amounts of data. DeepSeek reportedly aims to start pre-training workloads on the Ascend 950DT as soon as late 2026.

Equipment failures are exactly what tripped up earlier efforts. DeepSeek has reportedly struggled to train models on prior-generation Ascend chips because of stability and performance issues.

To avoid a repeat, Huawei engineers have reportedly embedded with DeepSeek’s team. The goal is to shape the chips’ design around the lab’s upcoming model requirements, rather than forcing the models to adapt to the hardware after the fact.

The hardware and the software behind it

On paper, the Ascend 950DT is built for both training and inference. It carries 144 GB of HBM-style memory and offers around 4 TB/s of bandwidth.

Huawei is targeting a commercial release of the chip in Q4 2026. That puts DeepSeek’s reported late-2026 pre-training timeline right on the heels of general availability, which leaves little margin for delays.

DeepSeek has also been working the software side. The lab has open-sourced several libraries optimized for Ascend hardware, which reportedly reached over 95% of the hardware’s peak performance on the Ascend 950 during tests.

Still, DeepSeek is not pretending the chips are a one-for-one Nvidia replacement. Founder Liang Wenfeng has said it takes four Ascend 950 chips to match the functionality of a single Nvidia GB300 GPU.

Why the pivot is happening now

The backdrop is US export restrictions on advanced AI chips bound for China. Those controls have pushed Chinese AI developers to look harder at domestic options, whether they wanted to or not.

DeepSeek’s approach pairs open-source software with high hardware utilization, a combination aimed at domestic hardware independence.

The broader roadmap points to mainstream adoption of Huawei’s Ascend stack as early as 2027.

What this means for Nvidia, Huawei and the AI chip race

Liang’s own four-to-one comparison suggests that running on Ascend may demand more chips, more power and more data center space for comparable results. That is a cost penalty DeepSeek is apparently willing to absorb for supply security.

The lab’s reported trouble with earlier Ascend generations shows that training stability is not guaranteed. The embedded Huawei engineers and the open-sourced libraries are designed to close that gap, but late 2026 will be the actual exam.

There is also a timing squeeze. A Q4 2026 commercial release and a late 2026 pre-training start overlap tightly. Any slip in Huawei’s production schedule could push DeepSeek’s V5 training timeline back with it.