DeepSeek wants to train an 8 trillion-parameter AI on Chinese chips – and its founder says it “must succeed”

DeepSeek is betting its future on Huawei. The Chinese AI lab’s famously low-profile founder, Liang Wenfeng, told investors at a closed-door meeting that switching to Chinese-made chips to train its AI models is “one of our biggest bets,” and that it “must succeed,” The Information reports.

The plan

DeepSeek plans to train models at scale on Huawei’s upcoming Ascend 950 chips, starting as early as late 2026. The goal is to reduce its reliance on Nvidia, whose most advanced chips China can’t legally buy because of U.S. export controls.

And the models it wants to train are enormous. DeepSeek is already training a model with 2 trillion parameters, and is planning one with 8 trillion, according to Android Headlines. That would dwarf Alibaba’s Qwen3.8-Max, at 2.4 trillion, and Moonshot’s Kimi K3, at 2.8 trillion.

The catch: you need a lot more of them

Huawei’s chips aren’t as powerful as Nvidia’s, so DeepSeek will need far more of them. Training at this scale would take roughly 50,000 of Nvidia’s top GB300 chips, or as many as 200,000 of Huawei’s Ascend 950s, BigGo Finance reports.

Supply is a problem too. Huawei is short of memory and other components, and the new training chips aren’t expected to arrive until the fourth quarter of this year or early next. So for now, DeepSeek is running on two tracks, continuing to train on the Nvidia chips it already has while it brings Huawei’s hardware online.

The money

DeepSeek is spending around 30 billion yuan, roughly $4.5 billion, to expand its computing infrastructure, and is finalising a funding round that would value it at about 500 billion yuan, or roughly $74.6 billion.

Why it matters

DeepSeek shocked the AI industry in early 2025 by building a competitive model at a fraction of what U.S. labs were spending. If it can now train frontier-scale models on Chinese chips, it would be the clearest sign yet that U.S. export controls aren’t stopping China’s AI progress, only forcing it to find another route.

Liang’s own words suggest he knows how much is riding on it. “Must succeed” isn’t how you describe a sure thing.

Sources: The Information, Android Headlines, BigGo Finance