Daily AI Catchup
ModelDeepseekHuaweiChipsChinaNvidia

DeepSeek to train next model on Huawei chips

DeepSeek reportedly plans to pretrain its next model, V5, on Huawei's Ascend AI chips, which have mostly been used for inference so far rather than full training runs. A successful run would be a significant proof point that Chinese-made chips can handle frontier-scale training, not just serving models. It would also strengthen Huawei's position as a domestic alternative to Nvidia.

Why it matters

๐Ÿ’ป Developer ยท If Ascend chips prove capable of full pretraining, expect more non-Nvidia training stacks and tooling to mature, which could diversify your hardware options over time.

๐Ÿ“ฆ Product ยท This affects the broader AI supply chain story: more viable chip options eventually mean more competition and potentially lower compute costs.

๐ŸŽจ Design ยท Limited direct relevance, but cheaper compute alternatives could eventually mean more affordable AI-powered design tools.

๐Ÿ“ˆ Business ยท A geopolitical signal worth watching: Chinese AI independence from Nvidia hardware would reshape chip export leverage and competitive dynamics for US chipmakers.

๐Ÿค” Just Curious ยท China is trying to build a major AI model without relying on Nvidia chips at all, a big test of whether the US chip restrictions are actually working.

Sources: DeepSeek Plans to Train Its Next-Generation Model on Huawei Ascend Chips