According to The Information, Liang Wenfeng told investors DeepSeek plans to use Huawei chips for training, potentially 950DTs arriving in Q4 2026 or Q1 2027.
He is reportedly training a 2T-parameter model and planning an 8T model, up from V4’s 1.4T.
Liang estimated a GPT-like model could require 50,000 Nvidia GB300s or 200,000 Huawei 950DTs. Limited Huawei chip and HBM supply means DeepSeek still needs more Nvidia B-series capacity.