Article
What Huang Renxun is most worried about is happening:DeepSeek the beginning of a move away from the US-led AI technology stack, with the underlying components collectively being adapted for Huawei Ascend
Beating AI breaking news reported that DeepSeek joined forces with Huawei to move an entire set of AI underlying components for training and inference onto Ascend. The newly open-sourced tools cover operator development, matrix computation, cross-card communication,Attention and data filtering, corresponding broadly to DeepSeek the underlying toolset previously built by Nvidia for GPU. Reuters called this the latest development in Chinese technology companies’ search for alternatives to Nvidia’s ecosystem.
The key point here is TileLang.DeepSeek A large number of operators in V4 training have already been implemented with it. It addresses a very fundamental issue: how developers can run the computations in a model efficiently on a chip. In the past, this capability was heavily built around CUDA; now DeepSeek and Huawei are providing the corresponding version for Ascend.
This directly hits at what Huang Renxun was most worried about several months ago. This year,April he said on Dwarkesh Patel’s podcast that if 1 day DeepSeek were released first on Huawei chips, it would be a “bad outcome” for the United States. What he has always worried about is not just Huawei’s chips themselves, but models beginning to be optimized around Huawei’s architecture. Once developers worldwide receive the same model and it runs better by default on non-US hardware, the advantage of US chips will gradually erode.
DeepSeek’s subsequent moves have followed this path almost exactly. V4 has already begun adapting to Ascend 950, and Huawei has also said that its chips participated in part of V4’s training. The Information reported last week that Liang Wenfeng had made increasing training on domestic chips a key priority for DeepSeek and expected to receive new Huawei training chips as early as the fourth quarter.
What is being added now is the software layer. From model adaptation and training chips to matrix computation,MoE communication,Attention and operator development,DeepSeek is gradually moving some of the underlying capabilities that previously relied heavily on Nvidia and CUDA to Huawei Ascend.
The US-China AI competition used to be about whose models were stronger and who had more GPU; now it is becoming a question of who can build an entire technology stack spanning chips, operators, communications, and training frameworks.CUDA What has always been hardest to replace is not a graphics card, but the complete software ecosystem that grew around it.DeepSeek What