Certain modified A100 GPUs are available at a price range from 80,000 yuan to 90,000 yuan, according to the Huaqiangbei vendor who declined to be identified.
The vendor, however, indicated that it would be “very difficult, if not impossible” to get a set of eight original A100 GPUs for a high-performance server system.
The A100 Tensor Core GPU, according to the Nvidia website, powers the world’s leading data centres used for AI, data analytics and high-performance computing applications.
Nvidia Corp’s H100 Tensor Core graphics processing unit is used to speed up large language models that are behind conversational artificial intelligence applications. Photo: Handout
On e-commerce giant JD.com’s platform, an authorised reseller of Nvidia listed the company’s 80-gigabyte A800 GPU for 88,999 yuan, while an 80G A100 was priced at 99,999 yuan, but was out of stock.
Nvidia’s A800 is allowed for export to China because it reduced some of the capabilities of the original A100. A tweaked China-export version of the H100 is called the H800.
The high local demand for smuggled GPUs underscores how China lacks strong alternative suppliers that can deliver products to rival those from Nvidia.
Still, the country’s major tech companies remain focused on getting Nvidia products. TikTok and Douyin owner ByteDance, for example, reportedly ordered US$1 billion worth of GPUs from Nvidia this year.
Reuters last week first reported about China’s thriving underground market for Nvidia GPUs.
Tang, the Shanghai-based intermediary for smuggled Nvidia GPUs, said the A100 can now easily sell for between 130,000 yuan and 150,000 yuan in the Yangtze River Delta, where many AI start-ups operate.
“Training an AI system is just a start,” Tang said. “If companies want to provide large language model-based services, they will require even more computing power.”
Large language models (LLMs) are deep-learning AI algorithms that can recognise, summarise, translate, predict and generate content using very large data sets. These represent the technology used to train AI chatbots like Microsoft-backed OpenAI’s ChatGPT.
Nvidia’s A800 chips, according to Tang, have been bought in bulk by several Chinese Big Tech companies over the past few months, which has led to a shortage in the local market. That situation, in turn, has further pushed up the prices of A100 and H100 GPUs bought from unofficial channels.
Tang said two of his customers needed to pay more than 1 million yuan upfront to lock in their supply for the next shipment of A800 GPUs. “First paid, first served,” he said.
China’s lack of access to advanced chips from the likes of AMD and Nvidia is already expected to hinder the country’s progress in various cutting-edge applications, from LLM development to autonomous driving.
At present, Nvidia’s H100, A100 and A800 GPUs as well as AMD’s MI250 and MI250X series are considered the mainstream products used in AI-related computing applications, according to a recent report by market research firm TrendForce. It said Nvidia now has an 80 per cent market share in server GPUs worldwide.
Development of Baidu’s ChatGPT-alternative Ernie Bot, which was originally created on servers using Nvidia’s A100 GPUs, has switched to the A800 owing to US export controls, according to TrendForce. It projected Baidu’s demand for AI servers, for Ernie Bot and its other applications, to reach around 2,000 units this year.
As such, demand for Nvidia’s China-tailored A800 and H800 GPUs has pushed up the prices of these products “day by day”, according to a sales manager, who requested not to be named, at certified Nvidia solutions provider Sitonholy Technology.
