DeepSeek's Liang Wenfeng Says Training Future Models on Huawei Chips Is Company's Biggest Bet

DeepSeek founder and CEO Liang Wenfeng told investors at a closed-door meeting that training future models on Huawei and other domestic AI accelerators is 'one of the company's biggest bets' and 'must succeed,' according to The Information, in reporting relayed by DigiTimes and Pandaily on September 22. Liang said DeepSeek expects a fresh batch of Huawei Ascend 950-series training chips to begin arriving as early as Q4 2026 as the company trains a roughly 2-trillion-parameter model — about 1.5x the 1.4-trillion-parameter V4 — with an eventual 8-trillion-parameter model also planned. Liang estimated that matching an OpenAI-class training run would take roughly 50,000 Nvidia GB300 chips versus around 200,000 Huawei Ascend 950 chips, a gap he said reflects a per-chip efficiency disadvantage domestic supply and cost advantages can offset over time. The remarks mark a shift from DeepSeek's earlier Nvidia-centric training stack after an earlier attempt to train on Huawei's Ascend 910C reportedly fell short, and follow DeepSeek's September 4 order of 160,000+ Ascend 950DT chips for a new data center in Ulanqab, Inner Mongolia. Several social media posts describing the closed-door meeting were reported removed shortly after they circulated.