DeepSeek plans to order at least 160,000 of Huawei's Ascend 950DT chips to power a new data center in Inner Mongolia. That's according to Bloomberg's reporting, cited by tech-ish and The News. The chips would run DeepSeek's models for end users, not train new ones.
Where the chips are going, and why there
The planned site sits in Ulanqab, Inner Mongolia, roughly 350 kilometers northwest of Beijing, per tech-ish's report. The location is not incidental. It offers cheap renewable power and an average temperature of 4.3°C, which cuts the cost of cooling a facility packed with accelerator chips.
Inference now, training still on Nvidia
DeepSeek's plan keeps a clear split. Huawei's chips would run inference, meaning they answer live user requests, while DeepSeek keeps training new models on the Nvidia hardware it already has, according to tech-ish's reporting. That split matters: it means this order is not yet a claim that Huawei's chips can fully replace Nvidia for DeepSeek's most demanding workload.
A chip Huawei can't yet fully supply
The Ascend 950DT itself doesn't launch until the fourth quarter of 2026. Even once it ships, Huawei faces a memory shortage that will cap its 2026 output at the low hundreds of thousands of units, per tech-ish. Filling an order this size could take Huawei more than a year.
The data center itself reflects that timeline: it's expected to only partially come online in late 2027 or early 2028.
What this means for developers
This is a hardware story with a software consequence. If DeepSeek's inference stack increasingly runs on Huawei's Ascend chips instead of Nvidia's, that matters for tooling built around CUDA. It needs a parallel path for Huawei's CANN software stack to keep working there.
Watch this as part of a broader pattern, not an isolated deal. Export restrictions have locked Nvidia out of China's advanced-chip market, and Chinese AI labs are visibly building around that gap rather than waiting for it to close. Expect more announcements like this one from other Chinese labs over the coming year.
Don't read this as proof Huawei chips have caught up to Nvidia for training. DeepSeek's own plan keeps training on Nvidia hardware for now, which is itself evidence that the gap for the hardest workloads hasn't closed yet.
This article was first published on Tech AI Wire.
Also available in
Deutsch · 日本語 · Français · Español · Português
Top comments (0)