26

2023

-

06

Crazy looting of AI chips: dismantling old goods in the computer room

Since the beginning of this year, big models and AI 2.0 have become both hot dreams of wealth creation and increasingly an expensive "power game". Among the various thresholds constructed by technology, capital and talent, there is also an important entrance ticket - do you have "access" to get enough Nvidia high-end graphics cards.


01. Sell chips like Hermès
Since the beginning of this year, big models and AI 2.0 have become both hot dreams of wealth creation and increasingly an expensive "power game". Among the various thresholds constructed by technology, capital and talent, there is also an important entrance ticket - do you have "access" to get enough Nvidia high-end graphics cards.
Wang Fei is the sales manager of a server customization enterprise in East China. He described to the "market" how popular Nvidia graphics cards are: "The demand is so hot that some people take the A100 80GB purchased before the chip embargo policy came into effect from the computer room and resell it." These old chips not only have no warranty, but also have a price of more than 100000 yuan compared with the unit price of more than 80000 yuan around August last year.
Even though the old goods have been dismantled and fully utilized, the supply of graphics cards is still severely insufficient. The A100 40GB, which has relatively small memory and low training efficiency, is not popular even if it has a brand new one. More inquiries, less transactions
Specifically, these popular video cards refer to Nvidia's "Tesla" series products. When training AI large models, GPUs (graphics processors, also known as display chips) that are more efficient than traditional CPUs, especially high-end GPUs, are difficult to find.
In 2020 and 2022, Nvidia successively launched A100 chip and H100 chip, among which A100 graphics card provides 80GB and 40GB video memory versions, and H100 provides 80GB video memory. The larger the graphics memory, the larger the neural network it can run
From September 2022, the United States banned Nvidia and other companies from exporting high-end GPU chips to China. In response to the embargo policy, Nvidia launched the A800 and H800 graphics cards for China in 2022 and 2023 respectively.
However, nowadays, A800 chips within the policy allowed range have also experienced prolonged supply cycles and price increases in the market. Wang Fei revealed, "Before the big model became popular, the A800 would arrive in about two weeks. But now I conservatively estimate with the client that it will take within 8 weeks, even in an optimistic situation, it will take 4-6 weeks

Key words:News