According to the Ministry of Industry and Information Technology, the scale of China's AI industry will reach 500 billion yuan in 2023, with more than 4,400 AI companies, and by 2030, industrial substitution will exceed 1 trillion yuan, and by 2035, it is expected to reach 2 trillion yuan.
But behind such a big AI industry, it was basically Nvidia's chips that were relied on before, and the data for the first half of 2023 shows that in China's AI chip market, Nvidia chips occupy about 90% of the market share.
However, everyone also knows that because of the export control of the United States, NVIDIA's chip procurement has become difficult, and what can be bought is already a special version of the chip that has been castrated over and over again, and the ** chain is facing great hidden dangers.
Therefore, Japan** recently said that China's artificial intelligence (AI)-related large enterprises are currently stepping up to switch AI chips to Chinese products to avoid their own business being affected.
So who can replace Nvidia? At present, there are not many Chinese AI chip manufacturers that have surfaced, the first is Huawei, Ascend AI chips, which should be one of the most popular AI chips in China, plus Huawei also has an ecosystem similar to CUDA, so orders have also increased dramatically.
After Huawei, there is also Haiguang, starting from the fall of 2023, the new AI chip "Deep Computing No. 2" can also partially replace NVIDIA's AI chips.
In addition to Huawei and Haiguang, the emerging Moore Threads, Avoidance, etc., are all gearing up, wanting to take a share of the AI chip market and become one of the alternatives to NVIDIA.
But to be honest, it is not easy for these companies to replace Nvidia.
On the one hand, NVIDIA can use all of TSMC's processes, including the most advanced 3nm, but some domestic manufacturers, because they have been sanctioned, cannot use TSMC's advanced technology, and can only use more backward processes, in this case, the performance will be relatively poor.
Secondly, NVIDIA's CUDA ecosystem is currently difficult to replace, because most of the AI is trained based on CUDA, using other chips, which needs to be recompiled, which is a very large workload, and may not be able to be recompiled smoothly, because the ecology is different.
However, whether it can be replaced or not, domestic chips must be able to top, if the localization of AI chips stagnates, the technological gap with the United States and other countries will widen, and the performance of AI will also lag behind.