On May 7, reports surfaced indicating that DeepSeek has recently officially launched and open-sourced its latest-generation model, DeepSeekV4, while also announcing that the model has undergone comprehensive adaptation for Huawei's Ascend chips. This milestone facilitates the complete deployment of domestic large-scale models on domestic chips, encompassing both training and inference phases. It underscores the remarkable strides China's AI industry has made toward achieving self-sufficiency and control.

Amidst an increasingly stringent external technological blockade, the adoption of domestic chips has transitioned from a "contingency plan" to an "unavoidable necessity." The seamless integration of DeepSeekV4 with Huawei's Ascend chips not only presents a viable alternative to CUDA but also dismantles the long-held prejudice that "domestic computing power is incapable of supporting cutting-edge models." It demonstrates that, through the collaborative optimization of algorithms and hardware, domestic chips can fully cater to the inference requirements of large models with trillion-level parameters, thereby charting a new course for global AI development.

The pivotal significance of this adaptation breakthrough lies in propelling the industrial structure to evolve from "isolated breakthroughs" to "systemic synergy." In addition to Huawei's Ascend, domestic chip brands such as Cambricon, Moore Threads, and Hygon have also declared immediate compatibility with DeepSeekV4, enabling the model to be deployed promptly upon release, without the need for additional debugging phases. This signifies the expedited formation of a collaborative ecosystem spanning the entire domestic computing power industry chain, from upstream to downstream.