China’s AI laboratory DeepSeek, whose AI types wreaked havoc on Wall Street last week, may use Chinese GPUs for future models, suggests a report. As investors began to wonder whether billions of dollars in capital expenditures were necessary for AI growth, DeepSeek’s ability to apparently reduce costs significantly over American types while using GPUs from chip big NVIDIA Corporation upended areas. Instead of NVIDIA’s offered language, Compute Unified Device Architecture ( CUDA ), the model development firm’s ability to use low-level programming language called Parallel Thread Execution ( PTX ) on NVIDIA’s GPUs is at the heart of the company’s model development.
According to a report, DeepSeek Could Rely On PTX Expertise to Get Higher Performance Out Of Domestic Chinese GPUs.
Chinese semiconductor manufacturing has frequently been the target of US sanctions, which have prevented its largest foundry, the Semiconductor Manufacturing International Corporation ( SMIC ), from acquiring the most recent generation of chips from Dutch powerhouse ASML. These tools are essential for producing the most cutting-edge chips, with several reports suggesting that SMIC is only allowed to produce chips using the more recent 7-nanometer process technology family.
The company’s ability to use a programming language called Parallel Thread Execution or PTX was a key factor in DeepSeek’s ability to match the performance of some Western AI models with the performance of less and less effective NVIDIA GPUs. There are comparatively fewer engineers who are capable of using PTX, which allows engineers to control a chip’s basic functions more deeply than other languages.

Due to US GPU sanctions, DeepSeek’s decision to use the notoriously hard PTX resulted from the company’s ability to secure either limited quantities of GPUs or only those with lower performance. According to a from the Hong Kong press, the company’s expertise in using PTX may allow it to rely on domestic Chinese GPUs for future models.
The report claims that because PTX enables programmers to rigorously regulate a GPU’s operations, it will help DeepSeek’s engineers extract more performance from domestic Chinese GPUs. The 5-nanometer manufacturing process node is used to create the Hopper H800 GPUs from NVIDIA, according to DeepSeek. They were established in 2023 and had at the time employed the most recent manufacturing process technology in the sector.
DeepSeek can use similar programs with Chinese GPUs by understanding NVIDIA’s PTX implementation and how drivers manage a chip’s functions, according to the report. Any Chinese GPUs that are built using older manufacturing process technologies will likely be significantly less performant than NVIDIA’s most recent Blackwell GPUs.