NVIDIA( NVDA) has requested TSMC to expand production in response to the surge in H200 GPU demand. Especially with Chinese customer orders surpassing 2 million units, this production adjustment is likely to bring NVIDIA huge profits. The ex-factory price of a single H200 GPU is approximately $27,000 (about 38.88 million KRW), and the total current order volume alone is expected to generate $54 billion (about 77.76 trillion KRW) in sales.
The H200 is an upgraded version of the GPU released in 2022. Its hardware architecture largely continues the design of the previous generation, but with significant improvements in memory capacity and bandwidth. Notably, high-bandwidth memory capacity has increased by about two times, and data transfer speed between memory and computing circuits has improved by nearly 40%. This enables it to achieve twice the computational speed of previous products in AI inference tasks such as large language models.
The performance of the H200 is more than six times that of NVIDIA’s current top-tier GPU sold in China, the H20. The H20 was introduced as an alternative after the U.S. government banned exports of top-tier GPU products to China, and the H200 was initially included in the export control list. However, the U.S. government later adjusted its stance, allowing exports of the product to China while requiring NVIDIA to pay a certain percentage of export sales to the government.
The initial profit remittance ratio was 15%, but it has recently been increased to 25% following directives from President Trump. President Trump also approved the sale of H200 to certain certified Chinese companies. This is seen as a compromise measure by the U.S. government to maintain technological controls while partially safeguarding NVIDIA’s commercial interests.
NVIDIA currently holds about 600,000 H200 chips and 100,000 GH200 “Grace Hopper” superchips in inventory. The GH200 is a high-performance chip combining two H200 GPUs with a 72-core CPU, and these inventories are also planned to be supplied to Chinese customers. The company intends to use existing stock to fulfill the initial order of 700,000 units before February 2026, with the remaining 1.3 million units to be supplied through increased production by TSMC in the second half of the year.
However, whether the expansion plan can proceed as scheduled remains uncertain. Even though the U.S. government has approved the export, Chinese import licenses have not yet been granted. Nevertheless, NVIDIA appears to have decided to start expanding production early to proactively meet customer demand.
Leading Chinese tech companies are driving bulk purchases, and even though the next-generation GPU “Blackwell Ultra,” which offers more than ten times the performance of the H200, has already been launched, demand for the H200 in the Chinese market remains strong. Blackwell Ultra provides over ten times the computational performance of the H200 and offers products that integrate two GPUs and one CPU as accelerators.
Through this initiative, NVIDIA is not only strengthening its supply chain of high-performance AI chips for China but also demonstrating a strategic posture of balancing global semiconductor policies and trade controls.
View Original
This page may contain third-party content, which is provided for information purposes only (not representations/warranties) and should not be considered as an endorsement of its views by Gate, nor as financial or professional advice. See Disclaimer for details.
China, Nvidia H200 2 million units sold out in an instant... Still racing ahead under Trump's "25% tariffs"
NVIDIA( NVDA) has requested TSMC to expand production in response to the surge in H200 GPU demand. Especially with Chinese customer orders surpassing 2 million units, this production adjustment is likely to bring NVIDIA huge profits. The ex-factory price of a single H200 GPU is approximately $27,000 (about 38.88 million KRW), and the total current order volume alone is expected to generate $54 billion (about 77.76 trillion KRW) in sales.
The H200 is an upgraded version of the GPU released in 2022. Its hardware architecture largely continues the design of the previous generation, but with significant improvements in memory capacity and bandwidth. Notably, high-bandwidth memory capacity has increased by about two times, and data transfer speed between memory and computing circuits has improved by nearly 40%. This enables it to achieve twice the computational speed of previous products in AI inference tasks such as large language models.
The performance of the H200 is more than six times that of NVIDIA’s current top-tier GPU sold in China, the H20. The H20 was introduced as an alternative after the U.S. government banned exports of top-tier GPU products to China, and the H200 was initially included in the export control list. However, the U.S. government later adjusted its stance, allowing exports of the product to China while requiring NVIDIA to pay a certain percentage of export sales to the government.
The initial profit remittance ratio was 15%, but it has recently been increased to 25% following directives from President Trump. President Trump also approved the sale of H200 to certain certified Chinese companies. This is seen as a compromise measure by the U.S. government to maintain technological controls while partially safeguarding NVIDIA’s commercial interests.
NVIDIA currently holds about 600,000 H200 chips and 100,000 GH200 “Grace Hopper” superchips in inventory. The GH200 is a high-performance chip combining two H200 GPUs with a 72-core CPU, and these inventories are also planned to be supplied to Chinese customers. The company intends to use existing stock to fulfill the initial order of 700,000 units before February 2026, with the remaining 1.3 million units to be supplied through increased production by TSMC in the second half of the year.
However, whether the expansion plan can proceed as scheduled remains uncertain. Even though the U.S. government has approved the export, Chinese import licenses have not yet been granted. Nevertheless, NVIDIA appears to have decided to start expanding production early to proactively meet customer demand.
Leading Chinese tech companies are driving bulk purchases, and even though the next-generation GPU “Blackwell Ultra,” which offers more than ten times the performance of the H200, has already been launched, demand for the H200 in the Chinese market remains strong. Blackwell Ultra provides over ten times the computational performance of the H200 and offers products that integrate two GPUs and one CPU as accelerators.
Through this initiative, NVIDIA is not only strengthening its supply chain of high-performance AI chips for China but also demonstrating a strategic posture of balancing global semiconductor policies and trade controls.