DeepSeek plans to pre-train its next-generation artificial intelligence model, V5, primarily on Huawei Ascend chips rather than Nvidia hardware, according to reports published on November 3 citing Chinese industry sources. Neither DeepSeek nor Huawei has confirmed the plan, and the reports stress that the project is still at the planning and selection stage.
The reports, carried by TrendForce citing Caiwen News and Star Market Daily, say V5 could become DeepSeek’s first large language model trained natively on Huawei’s Ascend 950 chips, with the project expected to materialise in 2027. A separate briefing citing the 21st Century Business Herald and STAR Market Daily describes the same plan: pre-training V5 on Huawei Ascend computing power.
If it goes ahead, the move would be a major test for China’s domestic AI hardware. Until now, Huawei’s Ascend chips at DeepSeek have been used mainly for inference — running finished models to generate responses — rather than for pre-training, the heavy first stage in which a model acquires its fundamental knowledge from scratch.
Why this would matter
Pre-training a frontier-scale model is the most compute-intensive part of building it, and it has so far been dominated by Nvidia’s chips and its CUDA software ecosystem. A successful full pre-training run on a wholly domestic Chinese hardware stack would show that a top Chinese lab can build a frontier model without Nvidia hardware — a question that has faced scepticism until now.
Reports say DeepSeek has already laid part of the groundwork. In late September 2026, it open-sourced its software stack for Ascend, including the TileLang kernel compiler, computational libraries and distributed communication libraries, mirroring its Nvidia components. WCCFTech reports that DeepSeek’s bespoke stack now runs natively on Ascend, lifting each chip’s Model FLOPS Utilisation to about 45%, up from around 30% previously reported for Huawei’s 910C chips.
The scale of the reported hardware build-out is large. WCCFTech and other reports say DeepSeek has ordered around 160,000 Ascend 950DT chips from Huawei for a 1-gigawatt data centre in Ulanqab, Inner Mongolia, with deliveries ramping up in the fourth quarter of 2026. Huawei’s Atlas 950 SuperPoD infrastructure links up to 8,192 Ascend 950DT chips to deliver 8 EFLOPS of peak AI compute, according to the reports. Some reports say large-scale pre-training could begin as early as late 2026, depending on testing progress, while the full project is expected in 2027.
What is not yet known
A great deal remains unconfirmed, and readers should treat the details as reported rather than established fact:
- V5 itself has not been officially announced by DeepSeek, and its specifications, capabilities and launch timeline are undisclosed. Industry sources expect a release sometime in 2027, hinging on training progress.
- The exact chip configuration, the number of chips a full pre-training run would need, and the timeline are all sourced to industry insiders, not to company statements. One report puts a complete pre-training run at more than 10,000 chips running for several weeks.
- Huawei holds no stake in DeepSeek, according to Caiwen News cited by TrendForce; cooperation is limited to products, computing resources and technical optimisation.
- The reports say Ascend remains DeepSeek’s primary training platform, with no reported adoption of other Chinese chipmakers’ solutions.
DeepSeek’s current production flagships are its V4 models, and V5 — if the reported plan holds — would follow them. The reports also note that Huawei has accelerated its own chip roadmap, with the Ascend 960DT now expected in the first quarter of 2027, three quarters ahead of its original schedule.
Why now
The reported shift comes as Chinese AI labs face uncertain access to Nvidia hardware, which remains subject to export controls and availability swings. Building a domestic training stack is therefore both a technical and a strategic goal. But the reports are clear that success cannot yet be guaranteed: training stability, cluster performance, software maturity and the resulting model’s capabilities are all still to be proven.
For now, the story is a reported plan, not a finished achievement. No user-facing DeepSeek model changes as a result of these reports: V5 does not yet exist, pre-training has not started, and the Ascend 950DT has not yet hit the market.
Frequently asked questions
Has DeepSeek confirmed it will train V5 on Huawei chips?
No. The plan is reported by Chinese outlets citing industry sources. Neither DeepSeek nor Huawei has confirmed it.
What would be new about training on Ascend chips?
Ascend chips at DeepSeek have so far been used mainly for inference. Pre-training a frontier model from scratch on them would be a first, and would test whether a domestic Chinese hardware stack can support the heaviest stage of model building.
When could V5 arrive?
Industry sources cited in the reports expect the project to materialise in 2027, with large-scale pre-training possibly beginning as early as late 2026, depending on testing progress.
Does this affect DeepSeek users today?
No. V5 does not exist yet and pre-training has not started, so the availability or quality of current DeepSeek models is unchanged.
Sources
- TrendForce, “DeepSeek V5 Could Complete Training on Huawei Ascend 950 in 2027, Moving Beyond Inference”, November 3, 2026: https://www.trendforce.com/news/2026/11/03/news-deepseek-v5-could-complete-training-on-huawei-ascend-950-in-2027-moving-beyond-inference/
- WCCFTech, “China Going For Domestic Full-Stack AI Dominance: DeepSeek Will Train Its Upcoming V5 Model From Scratch On Huawei’s Ascend 950DT-Based Atlas SuperPoDs”, November 3, 2026: https://wccftech.com/china-going-for-domestic-full-stack-ai-dominance-deepseek-will-train-its-upcoming-v5-model-from-scratch-on-huaweis-ascend-950dt-based-atlas-superpods/
- Kimkj China AI briefing, November 4, 2026: https://kimkj.com/en/ai-board-en/?mod=document&uid=20761
- MLLLM, “DeepSeek Prepares to Shift V5 Training from NVIDIA to Huawei Ascend”, November 3, 2026: https://www.mlllm.io/articles/3704-deepseek-prepares-to-shift-v5-training-from-nvidia-to-huawei-ascend-at-least-160-000-chips-to-be-deployed-in-ulanqab/
Related: Hg and Anthropic’s new European AI venture
