NVIDIA's new generation AI chip GB200 sees a surge in orders, while the H100 chip encounters a cold reception
The demand for NVIDIA's new generation AI chip GB200 has surged, while the market demand for the H100 chip has significantly decreased. The server price of the H100 chip has dropped from around 3 million RMB last year to about 2.3 million RMB, and the leasing price has also decreased from the peak of 8 USD/hour to 2-3 USD. With the continuous decline in domestic computing power prices, some small computing power suppliers may face losses, and the market is undergoing reshuffling
As the demand for NVIDIA's new generation AI chip GB200 surges, the once popular H100 chip is experiencing a bit of a cold spell. According to an industry insider who operates a computing power resource pool, servers equipped with the H100 chip are generally sold or rented out in an 8-card configuration. Last year, the selling price of an 8-card machine was generally above 3 million RMB, but now the price has dropped to around 2.3 million to 2.4 million RMB, with prices changing weekly. The price of an 8-card machine with the H20 chip, specially supplied by NVIDIA to China, has also dropped from around 1.4 million RMB to less than 1.1 million RMB. The initial rental price of the H100 chip was $4.7 per hour, but during the peak demand period, it was once speculated to be over $8 per hour. A search on a computing power rental website shows that the rental price for one H100NVL is now between $2 to $3 per hour, a decrease of over 50% compared to the peak period. According to industry insiders, since the first half of this year, the price of computing power in China has been on a downward trend. With a decline in demand for large-scale model training and an oversupply caused by early "hoarding" of cards, some smaller computing power suppliers may face losses and are being quickly phased out by the market
