DeepSeek plans to train V5 on 160,000 Huawei AI chips
DeepSeek reportedly plans to pre-train its next-generation V5 model on at least 160,000 Huawei Ascend 950DT chips at a data centre in Inner Mongolia, with work expected to begin in late 2026. TechNode, citing STAR Market Daily, says this would extend Huawei chips’ role from running DeepSeek models to training them; the plan has not been confirmed by DeepSeek or Huawei.
Bottom line — The V5 run will test whether Huawei’s chips and DeepSeek’s software can support large-scale model training.
Go deeper 6
-
Dataconomy reports the chips are intended for a new facility in Ulanqab, Inner Mongolia.
-
TrendForce says V5 remains in a selection phase, with its specifications and launch date undisclosed; sources it cites expect a release sometime in 2027.
-
Crypto Briefing reports that DeepSeek previously encountered stability and performance problems training on earlier Ascend chips.
-
Crypto Briefing says Huawei engineers have reportedly worked with DeepSeek’s team to adapt the chips to the lab’s requirements.
-
Wccftech reports that DeepSeek’s TileLang compiler and other software tools have been adapted for Huawei’s Ascend platform.
-
Geopolitechs notes that a Shenzhen research consortium completed post-training of DeepSeek V4-Pro on Ascend 910C chips; that was not pre-training from scratch.