07/10 2024 468
Foreign media data revealed that since 2023, despite repeated criticisms of low performance and high prices, NVIDIA's H20 chips, customized for the Chinese market, have seen significant purchases by domestic AI enterprises after nearly half a year of trialing domestic AI chips.
Since 2021, the United States has frequently restricted NVIDIA from selling high-performance AI chips to China. To comply with US export requirements, NVIDIA has introduced custom chips A800 and H800, which are castrated versions of A100 and H100, for the Chinese market.
In the second half of 2023, the US further blocked NVIDIA from selling high-end chips to China, including A800 and H800. NVIDIA had to further castrate performance and introduce the H20 chip. Foreign media pointed out that the performance of the H20 chip is only about 15% of high-end chips, but it still has powerful bandwidth and can achieve powerful performance by aggregating more H20 chips.
Due to NVIDIA's drastic performance castration, domestic AI chips found an opportunity. It is reported that the performance of a domestic AI chip is twice that of H20, while their prices are similar. Since last year, domestic AI enterprises have begun to try to replace NVIDIA's chips with this domestic AI chip, and there were even news that this domestic AI chip had received 1 million orders.
However, foreign media pointed out that these AI enterprises quickly discovered many problems. As this domestic AI chip only uses a process close to 7 nanometers, it has a large chip area and high power consumption, facing heating issues when implementing multi-chip integration. The bigger problem is the CUDA ecosystem. Global AI software has long been built on NVIDIA's CUDA platform, and adopting domestic AI chips requires redesigning software to achieve compatibility with the previous CUDA platform.
Foreign media pointed out that even so, over a million H20 chips have been sold in the Chinese market. Based on the H20 pricing of $12,000 to $16,000 per chip, NVIDIA has earned at least over $12 billion in the Chinese market, still achieving growth compared to previous years, although its revenue contribution from the Chinese market has dropped from a previous quarter to around 9%.
As domestic AI chips are used to build large models, these AI enterprises have discovered more problems. In actual use, domestic AI chips struggle to match the efficiency of NVIDIA chips. For AI enterprises, efficiency is paramount, especially as China's AI large models accelerate their pursuit of the United States, and such time costs are unbearable.
As a result, since the second quarter of this year, several domestic internet and autonomous driving companies have significantly increased their purchases of H20 chips. In particular, Tesla has once again changed its technical direction, using AI large models to retrain autonomous driving technology, proving that end-to-end autonomous driving is more efficient. Foreign media believes that NVIDIA's sales in China this year are expected to increase significantly, generating more revenue.
Regarding a certain domestic AI chip, foreign media pointed out that although its nominal performance is strong, its actual performance varies among internet companies, who were not entirely satisfied after initial trials. Ultimately, the number of domestic AI chips purchased by many domestic enterprises may only be 500,000, about half of the initial expectation.
Although the performance of domestic AI chips is somewhat below expectations, it still represents significant progress. The estimated 500,000 AI chips are expected to generate over $6 billion in revenue for them. In comparison, the fact that domestic AI chips have monopolized 90% of the market share from NVIDIA represents a huge step forward. As AI enterprises encounter issues with domestic AI chips, it can accelerate the technological upgrade of domestic AI chips, and it is believed that domestic AI chips still have a promising future in replacing NVIDIA.