Huatai Securities Research believes that DeepSeek uses PTX, which is a lower-level code than CUDA, to optimize hardware algorithms in V3. PTX is the intermediate code compiled by CUDA, which serves as a bridge between CUDA and the final machine code. NSA uses the Triton programming language proposed by OpenAl to efficiently write GPU code. The bottom layer of Triton can call CUDA, as well as other GPU languages, including AMD's rocm and domestic computing chip languages, such as Cambrian's Siyuan 590 chip and the HYGON ISA instruction set built into Haiguang Information's Deep Computing No. 1 (DCU). Although LLM training has not completely separated from the CUDA ecosystem in the short term, the launch of DeepSeek NSA has initially shown a trend of decoupling from CUDA and laid the foundation for subsequent adaptation to more types of computing chips. Domestic computing power represented by Yiteng has been well adapted to domestic models such as DeepSeek-R1 and has achieved efficient reasoning. Huatai Securities believes that with the limitation of overseas computing power, the optimization of domestic computing power may continue to progress and deserves attention.