大模型网摘Report Says DeepSeek V5 Pre-Training Could Shift to Ascend 950
Where the report comes from
On November 3, China's STCN Daily and Caiwen news, both citing supply-chain sources, reported that DeepSeek plans to use Huawei's Ascend 950 as the core training hardware for its next-generation model, V5. Local tech media framed it as a notable step for domestic compute in frontier-scale pre-training. As of publication, neither DeepSeek nor Huawei has confirmed the plan.
Cross-referenced reports add numbers. The Emergent and Digitimes versions mention roughly 160,000 Ascend 950DT chips to be deployed at a new Ulanqab data center in Inner Mongolia, with pre-training targeted to start in late 2026 and a possible debut in 2027. All of this remains unverified supply-chain chatter, so read it with a degree of caution.
Why Ulanqab
The compute layout is the most solid part of the story. Multiple sources mention the Ulanqab intelligent computing center as an active project, with Ascend 950 capacity as the planned backbone. The location logic is familiar. Cheap power, a cold climate, and a site other cloud providers have already used for training clusters.
Moving frontier pre-training onto a relatively young domestic hardware stack is hard engineering. Public technical discussions state the risk plainly. A multi-month run on new silicon is vulnerable to loss instability, and rework costs more than the switch itself. That is why outside observers see this as a stress test for the domestic software ecosystem, not just a procurement story.
It follows the late-September open-source move
Zoom out and the timeline fits together. On September 30, DeepSeek released a full set of open-source components for the Ascend environment, mirroring the NVIDIA stack from kernels to parallel training frameworks. At the time it read as support for inference clusters. In hindsight it looks like groundwork for a training-side switch. See our earlier coverage of the Ascend component release.
Money is the other backdrop. Late-October reports put DeepSeek's new round at roughly 15 billion US dollars, with CATL and Tencent participating and an IPO on the agenda. More in our funding report.
What users can rely on today
Shipping models remain the V4.1-Flash-1017 generation, updated on October 17 with agent-focused training, plus Search and Fetch APIs that entered paid service in November. Even on the rumored schedule, the road from pre-training start to a usable model is long. Our notes on the V4.1-Flash-1017 release cover the current lineup.
Things worth tracking. First the Ulanqab delivery progress and the actual scale of 950DT deployment, then any official word on the training platform. Until DeepSeek confirms, treat the details as rumor. This article is compiled from publicly available online sources.