NVIDIA’s special edition AI chip, H20, specifically designed for the Chinese market, is now available for pre-orders. The products offered include individual computing cards and servers equipped with eight H20 computing cards. According to insiders quoted by Chinese media Jiemian, the servers carrying eight H20 cards command a wholesale price exceeding 1.5 million yuan. This price is viewed by many dealers as inflated, causing them to hesitate on whether to stock the product. Additionally, the price of H20 computing cards is not fixed and is subject to fluctuations within a certain range.
NVIDIA introduced the H20 as a “special supply” version to comply with new export control regulations from the United States, which has restricted the sale of more advanced AI chips to China. The H20’s performance is approximately one-sixth of that of the H100, but the price hasn’t seen a significant reduction, raising questions about its cost-effectiveness.
NVIDIA reportedly has plans to introduce two more chips, the L20 and L2, specifically for the Chinese market. However, these have yet to appear on NVIDIA’s official website. The H20 was revealed three months ago but its official release has been delayed due to challenges faced by server manufacturers during the chip integration process.
China represents NVIDIA’s third-largest market, contributing to a fifth of the company’s total revenue. NVIDIA’s products, providing foundational computing power for AI development, are nearly irreplaceable in large model training. Previous models specifically launched for China, such as the A800 and H800, were in high demand and quickly sold out.
However, the future sales of the H20 remain uncertain. Some domestic cloud manufacturers have indicated that they have already purchased sufficient AI accelerator cards, implying no immediate need for the H20. Despite this, Shanxi Securities maintains a positive outlook, noting that while the H20’s computing power is less than that of the H100 and A100, its architecture and CUDA ecosystem have distinct advantages. Furthermore, stacking 2-3 H20s can achieve the performance of a single A100 card, suggesting potential for the H20 in large-scale models and providing it with a promising market scope.