Morgan Stanley (Da Morgan), a foreign-funded company in the United States, issued the latest AI** chain concern report, saying that some of Nvidia's H100 orders have shifted to H200 and B100, and it is expected that H100 sales will be 400,000 units in 2024.
Da Mo pointed out that Nvidia's GPU module shipments will be about 1.8 million units in 2023, and it is expected to double in 2024 and may be greater than 4 million. With the launch of a new production line in Mexico, Nvidia's GPU carrier board shipments will also increase significantly in 2024.
As for Nvidia's Chinese version of H20, which entered the mass production stage not long ago, the monthly output of chips in the first half of the year was as high as 20-300,000 pieces, but because the performance was worse than that of H100 chips, Chinese super-large-scale enterprises hesitated to place orders.
In the AI semiconductor ** chain part, Apple released the multimodal large language model Ferret, which runs large language models (LLMs) on mobile phones with mobile DRAM capacityIn terms of ASIC, Intel's Habana shipments were initially concentrated in Chinese customers' 7nm Gaudi 2, and 5nm Gaudi 3 shipments depended on TSMC's CODOS-S production capacity.
After the U.S. restricted the export of AI GPUs from China, NIO launched its own design of 5nm ADAS chips, which are expected to be mass-produced from 2026 to 2025, and the ASIC cost is $700-800 per set, which Da Mo believes is in line with NIO's goal of reducing the cost of self-driving systems from $4,000 to $1,500, which is comparable to Tesla.
Header image**: Shutterstock).