CEO Liang Wenfeng hosted an investor meeting Sunday to outline the company's long term strategy for infrastructure. DeepSeek is investing 30 billion yuan to expand compute capacity and is currently training a model with 2 trillion parameters. This shift to domestic hardware is a "major priority" intended to eventually scale to 8 trillion parameters, far exceeding the 1.4 trillion used in its current V4 version.

Huawei deliveries of AI training chips arrive in Q4 2026 or Q1 2027. DeepSeek continues to train its newest AI models using Nvidia chips likely obtained via the black market while it joins forces with Huawei to bypass export controls. For context, the rival Kimi K3 model uses 2.8 trillion parameters.