DeepSeek plans to train AI on Huawei chips at large scale

By Splitfeed EditorialLast updated Sep 26, 2026, 20:18 UTC

DeepSeek, the Chinese AI lab founded by Liang Wenfeng, is planning to train its next major model on Huawei chips at large scale, a move Liang himself has described as the company's biggest bet and one that must succeed. The model in question is reported at 8 trillion parameters, which would place it among the largest ever attempted. New Huawei chip supply to support the training run is expected to arrive in the fourth quarter of 2026. Separately, DeepSeek has already confirmed that Huawei hardware is capable of handling training workloads, not just inference, which marks a meaningful shift in how those chips have been used up to this point.

The context that gives this story its weight is the U.S. export control regime that has progressively cut off Chinese AI labs from Nvidia hardware, which has long been the dominant choice for training large models. DeepSeek's commitment to Huawei chips at this scale is effectively a public test of whether domestic Chinese silicon can compete with Nvidia for the most demanding workloads in AI. Nvidia's China sales have been declining under the restrictions, and if DeepSeek's bet pays off, it would suggest the export controls are accelerating, rather than preventing, the development of a credible alternative supply chain. What remains genuinely in dispute is whether Huawei's chips can match Nvidia's performance and reliability at this scale, or whether the efficiency gap will impose serious costs on the resulting model.

The anti-AI camp has framed the pivot in skeptical terms. The account Zubiqo on Bluesky argued that accepting a reported four-to-one hardware disadvantage reveals the move as survival rather than strategic ambition, suggesting the shift is being driven by necessity rather than confidence in the underlying technology. The pro-AI camp has not yet published reactions, but a story like this would typically attract arguments that DeepSeek's track record of doing more with less hardware makes it a credible candidate to close the gap, and that a successful training run would validate the broader case for AI development outside Western supply chains. The middle ground camp has also not weighed in, though it would ordinarily focus on the genuine uncertainty around Huawei chip performance and the risks of drawing conclusions before the training run is complete.

The clearest thing to watch is whether new Huawei chip supply arrives on the timeline reported for Q4 2026 and, after that, what DeepSeek discloses about the performance and efficiency of any model trained on that hardware. Those results would go a long way toward settling the question of whether this bet succeeded.