China plans to approve the purchase of Nvidia H200 chips by scientific and technological enterprises!

The NVIDIA H200, as an upgraded version of the Hopper architecture, has become the key to alleviating the bottleneck of domestic AI computing power with its excellent performance. The chip is equipped with 141GB HBM3e memory, with a bandwidth of 4.8TB/s, which is 76% and 43% higher than the previous generation H100, respectively. It doubles the speed and halves the energy consumption in large model inference such as Llama 270B, and is compatible with existing H100 systems without the need for architecture reconstruction. For Alibaba, Tencent, ByteDance and other enterprises, H200 can accelerate the iteration of big models and the commercialization of generative AI.

Previously, due to the impact of US chip regulations, domestic companies had difficulty obtaining high-end AI chips, and the shortage of computing power constrained the development of the industry. The performance of the H20 supplied by Nvidia has shrunk, making it unable to meet the demand for billion parameter large models. The approval of H200 for import this time will inject computing power into the market. Analysts predict that Alibaba and Tencent's combined AI chip procurement expenses may increase by more than 4 billion US dollars next year, driving the domestic AI commercialization process to accelerate by 12 to 18 months.

The accompanying condition of "synchronous procurement of domestic chips" is a strategic layout that balances short-term demand and long-term security. On the one hand, importing H200 can alleviate the computing power shortage and ensure the global competitive pace of the AI industry; On the other hand, by leveraging import permissions, we aim to expand the market for domestic chips, accelerate technological and ecological maturity, avoid dependence on a single overseas supplier, and achieve a balance between efficiency and safety.

Currently, domestic AI chips are transitioning from being "usable" to "easy to use". Products such as Huawei Ascend and Cambricon continue to break through in performance and have been widely applied in the fields of government and finance. Policy guidance for purchasing domestically produced chips can provide market validation opportunities for local manufacturers and optimize the software and hardware ecosystem. This orientation is clear, and China's AI industry will not take the extreme route of "relying on imports" or "working behind closed doors".

The news is directly positive for Nvidia, as its pre-market stock price rise reflects the market's expectations for China's computing power demand. China is one of the fastest-growing markets for AI in the world. Previously, regulations led to a decline in Nvidia's high-end chip market share. The approval of the H200 is expected to help Nvidia recover some of its market share and alleviate performance pressure. However, it should be noted that the United States will attach 25% income handover conditions to H200, and the cost may be transferred to domestic enterprises, which will weaken its cost performance ratio.

For domestic chip companies, the introduction of H200 is both a challenge and an opportunity. Short term partial orders may flow back to Nvidia, putting pressure on companies such as Cambrian and Haiguang Information, causing fluctuations in their stock prices. But in the long run, it will form a 'catfish effect', forcing local enterprises to accelerate technological breakthroughs and narrow the gap between performance and ecology. Driven by policies and demand, domestic chips are expected to make breakthroughs in specific scenarios, forming a competitive landscape with overseas products.