Nvidia’s specialized AI chip, despite its reduced performance, could generate substantial revenue in China, potentially reaching $12 billion. On July 5, consulting firm SemiAnalysis released a report suggesting that Nvidia is expected to deliver one million H20 GPUs this year. These “special version for China” chips, which have been downgraded in performance, are expected to be sold exclusively in the Chinese mainland market. Priced between $12,000 and $13,000 each, these chips could bring in significant revenue, surpassing Nvidia’s total revenue in China from the previous year.
A server dealer informed Huxiu that an H20 eight-card server currently costs about 1.2 million yuan. In comparison, an eight-card server with the high-performing H100 NVLink version costs around 2.7 million yuan and is readily available. Although the H100 was included in the list of products banned for sale to China, it continues to be supplied through unofficial channels. Even after the H200’s release in the second quarter of this year, the H100’s price dropped by about 10%.
“No matter how you look at it, the price-performance ratio of H20 is still too low,” the dealer remarked, adding that even compared to domestic solutions, the H20 lacks competitive advantage.
A “Castrated Version” with Surprising Features
The H20, unlike other chips in its series, appears to have been hastily produced. Following the U.S. chip ban upgrade in October last year, Nvidia’s A800/H800 and consumer-grade RTX4090 were included in the ban. Yet, a month later, Taiwanese media reported Nvidia had ordered H20 chips.
Given the short timeframe, it’s unlikely Nvidia designed a new chip from scratch. Instead, they probably made targeted adjustments to existing models. For instance, Nvidia increased the HBM3 memory capacity of the H20 to 96Gb, raising the memory bandwidth to 4.0Tb/s, and cut the computing power specification to comply with new regulations. Despite these adjustments, the H20 retains features like PCIe Gen5 expansion cards and high inter-card interconnection speed, making it suitable for large-scale clusters.
However, these modifications come with a cost. The H20’s BOM is likely similar to or higher than the H100’s due to the high cost of storage particles, which constitute a significant portion of high-performance computing cards’ costs. Therefore, the H20 has the lowest profit margin among Nvidia’s high-performance computing cards.
Challenges and Market Dynamics
Nvidia needs the H20 to stabilize its market share in China after the chip ban led to a revenue drop. Despite being a conscientious product from Nvidia, the H20’s computing power is significantly lower, posing a challenge.
In February, Nvidia officially reduced the H20 chip’s price from $23,000 to $12,000-13,000 due to insufficient sales. Large domestic manufacturers prefer the H100, while smaller manufacturers find the H20 uneconomical. The rental cost for a single H100 server is about 80,000 to 90,000 yuan per month, while an H20 server, with only one-sixth the computing power, costs about 1.2 million yuan.
Domestic competitors like Huawei’s Ascend 910B also pose a threat. The 910B, though priced higher, offers almost twice the computing power of the H20. The H20, however, excels in software compatibility and interconnection, making it suitable for building computing clusters despite its lower computing power.
Supply Stability and Future Outlook
An industry insider suggested that for manufacturers in urgent need of building computing centers, the H20 might be a more practical choice due to Nvidia’s mature software ecosystem and quicker delivery compared to Huawei’s 910B. Server distributors confirmed this, noting increased interest in the H20 following Nvidia’s price adjustment.
Interestingly, the H20 and 910B may not be direct competitors, as some companies categorize them differently, avoiding procurement conflicts. This nuanced market dynamic could shape Nvidia’s strategy and revenue outcomes in China’s AI chip market.
Related topics:
Understanding Data: The Foundation of the Digital Ag
OpenAI Surpasses Microsoft in AI Model Sales with $1 Billion Annual Revenue