Training is the workload where Chinese labs still depend most on Nvidia; a DeepSeek V5 trained on Ascend would make Huawei a credible supplier for the whole model life cycle.
DeepSeek intends to pre-train its next-generation model, DeepSeek V5, on Huawei’s Ascend chips, STAR Market Daily (科创板日报) reported on November 3, citing several people in the industry supply chain. The report, republished by The Paper, says Ascend chips have until now been used mainly for DeepSeek’s inference, the work of answering users with a model that has already been trained. Neither DeepSeek nor Huawei has commented on the report, and DeepSeek has not announced V5.
The claim matters because of what it would move. Chinese chips now carry much of the country’s inference work, the 21st Century Business Herald noted, while the most demanding stage of model building has stayed harder to relocate. A full pre-training run would ask far more of the Ascend chips than anything DeepSeek has publicly used them for.
STAR Market Daily gave no date, model size or chip count. A separate Cailian Press report the same day, carried by Phoenix News and also citing industry sources, went further: V5 would use the Ascend 950 as its core training compute, and the project points to 2027, with the exact timing depending on how training progresses. Neither report says whether Ascend would carry all of pre-training or part of it alongside other hardware.
The partnership has moved from serving models toward their foundations
The relationship between the two companies has deepened in steps. The 21st Century Business Herald, carrying the STAR Market Daily report, traced large-scale Ascend use for DeepSeek models to the 2025 Lunar New Year. Citing a book on Huawei’s chip business, it said Huawei’s leadership decided on January 28, 2025 to support DeepSeek quickly, and that more than 3,000 Ascend chips were deployed and tested over the holiday to put DeepSeek-V3 and R1 online at Huawei Cloud’s Gui’an data center.
After DeepSeek released V4 in April 2026, Huawei said its full range of Ascend SuperPoD systems could support the V4 series. DeepSeek’s own documentation, as cited in the report, said a shortage of high-end compute was limiting how much V4-Pro traffic it could serve, and that V4-Pro prices should fall significantly once Ascend 950 SuperPoD systems ship in volume in the second half of the year.
The most concrete step came on September 30, when DeepSeek open-sourced a set of infrastructure software for Ascend: the TileLang compiler tools, computing libraries including DeepGEMM and FlashMLA, and the DeepEP communication library. Each matches a component DeepSeek had earlier released for Nvidia hardware. Huawei contributed jointly defined networking designs and opened its Ascend C programming interfaces to support the work. Software of this kind is what a lab needs before it can run a long, stable training job on new hardware.
Earlier reports described an inference cluster, not a training one
The new report sits uneasily with what Bloomberg described two months ago. In September, Bloomberg reported that DeepSeek planned to deploy at least 160,000 Huawei Ascend 950DT accelerators at a data center it is building in Inner Mongolia, with a timeline that depends on Huawei’s production capacity. According to TechNode, which summarized that report, the chips were intended for inference, and DeepSeek did not then plan to use them for training. EastFrontier covered that plan at the time: DeepSeek plans a giant Huawei chip cluster for inference.
The Cailian Press report sharpens the conflict, because it names the Ascend 950 family as V5’s training chip. It also offers one way to reconcile the accounts: DeepSeek’s planned computing center in Ulanqab, Inner Mongolia, is still at an early stage of construction, and the early training compute for V5 may not be located there, people familiar with the matter told Cailian. DeepSeek’s plans may also simply have changed since September.
The technical gap is real. In June, a team from the Shenzhen Loop Area Institute, the Harbin Institute of Technology’s Shenzhen campus and others, working with Huawei, said it had completed full-parameter post-training of the 1.6-trillion-parameter DeepSeek-V4-Pro on at least 1,000 Ascend 910C chips, Tom’s Hardware reported. Post-training adjusts a model that already exists. Pre-training builds it from scratch and demands far more compute and stability over a long run. The outlet wrote that domestic chips are still used mainly for inference and that the cost of moving training software to a new chip ecosystem remains a limiting factor for the market.
Ascend 950 shipments will decide how soon the plan can be tested
Any V5 training run on Ascend depends on Huawei delivering its Ascend 950 SuperPoD systems in volume, the same shipments DeepSeek has tied to lower V4-Pro prices, and Cailian’s 2027 timeline leaves room for that. If V5 does complete pre-training on Ascend, the 21st Century Business Herald wrote, it would give the first answer to whether a top Chinese model can be pre-trained on domestic compute.
For Huawei, that would turn DeepSeek from its most prominent inference customer into a reference customer for training, the segment where Nvidia’s position in China has been hardest to replace. For other Chinese labs, it would show whether the open-source Ascend software DeepSeek released in September is mature enough for frontier-scale work.
The clearest confirmation would come from DeepSeek itself: a V5 release, or a technical report, that names the hardware used for pre-training.