Nvidia is expected to re-enter the Chinese market it once lost. On December 9, according to reports from the Global Times citing Reuters, Bloomberg and other media, U.S. President Trump announced on social media on the 8th local time that the U.S. government would allow Nvidia to sell its H200 artificial intelligence chips to China, but would charge a 25% fee for each chip. He also said that the U.S. Department of Commerce is finalizing the details of relevant arrangements, and the same arrangements will also apply to other artificial intelligence chip companies such as AMD and Intel.
Today, a reporter from China Business News asked NVIDIA for confirmation of this news. NVIDIA responded: "Supplying H200 to commercial customers is a move worthy of recognition."
In July this year, when Huang Renxun came to China to be interviewed by China Business News and other media, he said that he looked forward to more advanced chips entering the Chinese market in the future. Now this expectation is expected to become a reality, although it comes with a "fee" condition.
The H200 chip will be released as early as November 2023 and will be shipped in 2024. Its performance is several times that of H20. Based on NVIDIA's "Hopper" architecture, the H200 is the first GPU to feature HBM3e (fifth generation high-bandwidth memory). It is designed for training and inferring today's generative AI models, running large-scale scientific calculations, and processing massive amounts of data. The H200 offers 141GB of memory at 4.8 TB per second, nearly twice the capacity and 2.4 times the bandwidth compared to the A100. Among them, the inference speed on Llama 2 (70 billion parameter LLM) is twice as fast as H100.
Nvidia’s previous financial report for the second quarter of fiscal year 2026 showed that Nvidia’s revenue for the quarter was US$46.743 billion, of which revenue from the Chinese market (excluding Taiwan, China) was US$2.769 billion. Compared with the Chinese market revenue of US$3.667 billion in the second quarter of fiscal year 2025, it shrank by nearly US$900 million.
In a public event in October this year, Nvidia CEO Huang Jenxun once said that U.S. policies have led to the loss of one of the largest markets in the world. NVIDIA's market share in China dropped from 95% to 0%, and NVIDIA left 100% of the Chinese market. Huang Renxun said that missing out on China's artificial intelligence (AI) market would be a "huge loss."
Huang Renxun also mentioned China's strength in the field of AI, saying that China is a unique market. The vitality, innovation ability, development momentum of this market, and the development speed of this industry are all unique. The unintended consequences and long-term effects of not participating in the Chinese market, although difficult to predict, will not be optimistic. He predicts that China's artificial intelligence market may reach about US$50 billion in the next two to three years. For an American company, not being able to enter this market would be a huge loss, and China's AI market will advance with or without Nvidia.
Next, factors such as when customers in the Chinese market place new orders, whether demand changes and how long the supply chain takes will affect Nvidia's performance forecast. More importantly, during Nvidia's "absence", the Chinese market has changed - the rise of local AI chip manufacturers has accelerated, and customers have found alternatives. In Tencent’s recent financial report conference, Tencent executives stated that the current GPU is fully sufficient and resources are not very tight. Baidu executives recently stated that the vast majority of reasoning tasks within Baidu currently run on its own Kunlun core P800. Zhou Hongyi, founder of 360 Group, also told reporters that chip procurement has now shifted to domestic chips.
At the same time, Nvidia continues to face competitive pressure from other self-developed AI chip manufacturers. Compared with the versatility of Nvidia GPUs, ASIC (application specific integrated circuit) chips represented by Google TPU have some advantages, such as higher theoretical energy consumption and performance. OpenAI competitor Anthropic has planned to use Google's computing power in October and plans to deploy up to 1 million Google TPU chips to train its large AI model Claude. This batch of TPU chips will be specifically used to accelerate machine learning workloads and is planned to be deployed in 2026.