DeepSeek V5 said to train on Ascend 950 chips in 2027

On November 3, 2026, rumors began circulating in Chinese tech and investment circles about the training setup for DeepSeek's next-generation large model, DeepSeek V5. The first report came from East Money at 14:27 Beijing time, citing multiple industry supply chain insiders who said DeepSeek V5 is planned to be pre-trained using Huawei Ascend chips, whereas Ascend had previously been used mainly for inference tasks on DeepSeek's models. The report noted that this training plan has not yet been confirmed by DeepSeek or Huawei. It also pointed to DeepSeek's own API documentation, which mentions that service throughput for V4-Pro remains limited due to high-end compute constraints, with further Ascend-related expansion expected in the second half of the year. Forty minutes later, the Reddit community LocalLLaMA picked up the same claim, referencing a report from STAR Market Daily that said the information came from multiple sources within the industry supply chain and was based on anonymous supply chain information, with neither DeepSeek nor Huawei publicly confirming it. At 16:08 the same day, cnBeta added further detail, citing supply chain insiders who said DeepSeek V5 is planned to use Huawei's Ascend 950 chip as its core training compute, with the project timeline pointing to a 2027 launch. According to this report, the core R&D coordination teams from DeepSeek and Huawei Ascend are stationed in Hangzhou, and the two sides have already completed substantial low-level operator adaptation and customized development during the V4 stage, laying groundwork for future model training migration and compute collaboration. In previous generations, Ascend hardware had been used for inference work but had not entered the training stage. All three reports emphasize that the information remains unconfirmed by the companies involved, and the current status is that the claim remains a rumor pending official verification.