DeepSeek reverses the impression that domestic chips are not easy to use: Huawei 950 super node can completely replace NVIDIA GB300 and others

Kuai Technology reported on August 22 that the recent internal speech of Liang Wenfeng, the head of DeepSeek, caused an uproar on the Internet. The calmness and confidence shown by this leader in the AI ​​industry in his speech, as well as his in-depth analysis of the entire industry, were not expected by many people who had followed him before. Compared with the practical issues such as commercialization and profitability of his own company that are generally concerned by the outside world, Liang Wenfeng’s current layout direction seems to be to use mature and iterative domestic chips with fully independent domestic AI large models, directly impacting the global AI technology narrative system that the United States has maintained for many years. Judging from the content of this circulated internal speech, Liang Wenfeng is more optimistic about the development prospects of domestic AI chips than most industry observers. He believes that the traditional moat of NVIDIA's CUDA is disintegrating faster than expected. There are three core reasons behind it: AI can already generate code independently to build a complete ecosystem. The newly launched various high-level programming languages ​​​​have greatly reduced the difficulty of rewriting operators. The market size of AI computing cards has completely exceeded that of game cards. NVIDIA has no reasonable reason to forcibly bind the ecology of these two product lines together. These technological changes, coupled with the realistic background of U.S. chip export controls, are equivalent to creating a historically unimaginable development window for the domestic AI chip industry out of thin air. He made a clear prediction that within the next year, the entire market will completely reverse the inherent biased perception that "domestic chips cannot be used at all." DeepSeek has now entered into in-depth cooperation with Huawei and obtained 16,000 Ascend 950 cards. Although the overall equivalent computing power is only equivalent to 4,000 NVIDIA B-series chips, which cannot yet support the training of the next-generation flagship large model, it is completely enough to help Huawei run through the verification of the entire computing power ecosystem. He also admitted that Huawei's current core problem has never been technical capabilities, but the ramp-up speed of production capacity. The comprehensive performance of four Huawei chips can roughly match that of an NVIDIA card. The paper parameters are two years behind. However, as long as the subsequent price and ecology can be equalized, this level of performance gap is completely acceptable to the entire industry. According to this industry leader, the outside world had previously worried about