What Jensen Huang fears most is happening: DeepSeek is beginning to move away from the US-led AI tech stack, with its underlying components collectively adapting to Huawei Ascend.
动察 Beating AI News Flash: DeepSeek has partnered with Huawei to migrate an entire set of AI training and inference low-level components onto Ascend. The newly open-sourced tools cover operator development, matrix computation, cross-card communication, Attention, and data filtering, essentially corresponding to the low-level toolset DeepSeek previously built for Nvidia GPUs. Reuters called this the latest progress by Chinese tech companies in seeking alternatives to the Nvidia ecosystem.
The most critical piece here is TileLang. A large number of operators in DeepSeek V4 training have already been implemented with it. It solves a very fundamental problem: how developers can actually run the computations in a model efficiently on the chip. In the past, this capability was built highly around CUDA, and now DeepSeek and Huawei are filling in the corresponding version for Ascend.
This hits precisely what Jensen Huang was most worried about a few months ago. In April this year, he said on Dwarkesh Patel's podcast that if one day DeepSeek releases first on Huawei chips, it would be a "terrible outcome" for the United States. What he has always worried about is not just Huawei's chips themselves, but models beginning to optimize around Huawei's architecture. Once global developers get the same model and it by default runs better on non-U.S. hardware, the advantage of U.S. chips will be eroded bit by bit.
DeepSeek's subsequent moves have unfolded almost exactly along this path. V4 has already begun adapting to Ascend 950, and Huawei also said its own chips participated in part of V4 training. The Information disclosed last week that Liang Wenfeng has made increasing domestic-chip training a key bet for DeepSeek and expects to receive new Huawei training chips as early as the fourth quarter.
What is being filled in now is the software layer. From model adaptation to training chips, and then to matrix computation, MoE communication, Attention, and operator development, DeepSeek is gradually migrating part of the low-level capabilities that previously depended heavily on Nvidia and CUDA onto Huawei Ascend.
In the past, the China-U.S. AI rivalry was about whose model was stronger and who had more GPUs; now it is beginning to become about who can build a complete technology stack from chips, operators, and communication to training frameworks. What has always been hardest to replace about CUDA was never a graphics card, but the entire software ecosystem that grew around it. What DeepSeek and Huawei are now filling in is precisely this layer.
