All
  • All
  • Product Management
  • News
  • Introduction
  • Corporate Outlets
  • FAQ
  • Corporate Video
  • Corporate Brochure
language

    Return to List

    AI chip frenzy: Data centers dismantle old equipment, selling individual chips for 100,000 yuan

    Jun 26,2023

    AI chip frenzy: Data centers dismantle old equipment, selling individual chips for 100,000 yuan

    Selling chips like Hermès bags

    Since the beginning of this year, large models and AI 2.0 have become both a hot wealth-creation dream and an increasingly expensive "game of power." Among the various thresholds built on technology, capital, and talent, there is another important entry ticket—whether you have the "connections" to obtain enough high-end Nvidia graphics cards.

    Wang Fei, sales manager of a server customization company in East China, described to "Shijie" the recent scarcity of Nvidia graphics cards: "Demand is too explosive. Some people have dismantled and resold the A100 80GB purchased before the chip embargo policy took effect from the server room." These old chips not only have no warranty, but their price has also risen to over 100,000 yuan, compared to the unit price of more than 80,000 yuan around August last year.

    Even with the resale of used and fully utilized goods, the supply of graphics cards is still severely insufficient. "You need to pay cash for them to be willing to sell to you." Even brand-new A100 40GB cards, with relatively smaller video memory and lower training efficiency, are unpopular. "Many inquiries, few transactions."

    Specifically, these sought-after graphics cards refer to Nvidia's "Tesla" series products. When training AI large models, GPUs (Graphics Processing Units, also known as display chips), which are more efficient than traditional CPUs, especially high-end GPUs, are in short supply.

    In 2020 and 2022, Nvidia launched the A100 and H100 chips, respectively. The A100 graphics card offers 80GB and 40GB video memory versions, while the H100 offers 80GB video memory. The larger the video memory, the larger the neural network that can be run.

    Starting in September 2022, the United States banned Nvidia and others from exporting high-end GPU chips to China. To cope with the embargo policy, Nvidia launched the A800 and H800 graphics cards specifically for the Chinese market in 2022 and 2023, respectively.

    However, now, even the A800 chips, which are within the permitted scope of the policy, are experiencing longer supply cycles and price increases in the market. Wang Fei revealed: "Before the large model boom, A800 could be delivered in about two weeks. But now, I conservatively estimate 8 weeks for my customers, and even in an optimistic scenario, it would take 4-6 weeks."