DeepSeek has never been shy about working around US chip restrictions, but what’s happening now is a different order of magnitude. According to Huawei Central, the Chinese AI startup is moving toward broad adoption of Huawei chips for AI model training, explicitly pulling back from Nvidia as its primary hardware foundation. This isn’t a pilot program or a hedge. It looks like a strategic commitment.
The timing matters. US export controls have progressively cut off Chinese companies from Nvidia’s most powerful datacenter GPUs, including the H100 and its successors. DeepSeek already made headlines earlier this year by training competitive models on older, restricted hardware. But shifting to Huawei’s Ascend chips at scale is a different move entirely. It means DeepSeek is no longer just working around restrictions, it’s actively building a supply chain that doesn’t depend on American silicon at all.
For Huawei, this is significant validation. The company’s Ascend 910B and 910C chips have struggled to gain serious traction outside of state-directed deployments, partly because of performance gaps compared to Nvidia’s A100 and H100 generation, and partly because the software ecosystem around them is still maturing. Having DeepSeek, one of the most credible AI labs in China right now, run large-scale training workloads on Ascend hardware gives Huawei something it badly needs: a real-world proof point.
The broader industry context is worth understanding here. China is not the only place watching this. Any major AI lab that can demonstrate high-quality model training on non-Nvidia hardware weakens the assumption that Nvidia’s stack is essentially mandatory. Competitors like AMD, with its MI300X, and startups like Cerebras and Groq are all trying to crack the same problem in Western markets. DeepSeek doing it with Huawei, at scale, adds pressure to that narrative from an unexpected direction.
What remains unclear is how much of a performance tradeoff DeepSeek is accepting, and whether the models trained on Huawei hardware will be competitive with those trained on the latest Nvidia infrastructure. DeepSeek has shown before that efficiency and clever architecture choices can compensate for raw hardware limitations. But this is a larger bet, and the results will be watched closely by anyone who thinks the AI hardware market is still wide open.
