Certain AI data centers in China have significant stocks of NVIDIA RTX 4090D GPU specific to China, to dismantle them and renovate them, then put them back on the market as new cards. Digitimes Asia Reports according to which some companies use the abandonment of the excess computer capacity to generate the benefit several times compared to the rental of the GPUs, which will lead them about three to five years to recover their investment.
This decision is a sign of the way the rush towards Chinese AI leads to Billions of dollars in inactive infrastructure. The report indicates that an AI data center requires a use rate of more than 70% to 75% to make a profit. However, activation rates remain less than 20%, which means that a significant amount of capacity is not used and many GPUs remain inactive. To help them stay afloat, a few companies turn to the sale of their unused assets to generate fast money and reimburse the bank loans they have used to buy their equipment.
The sale of RTX 4090 cards is also quite lucrative. Currently, these RTX 4090D GPUs specific to China with 48 GB of VRAM are at the price between 20,000 CNY and 40,000 CNY, or around US $ 2,735 and US $ 5,470. We do not know if these cards have been used, but they must always be modified if the data center wishes to sell them to consumers. Data centers generally convert GPUs cooled by fan into blowing style cards, like this RTX 5090D in blowing style fled to Bilibili, For improved efficiency when used in multi-GPU systems. However, these are much noisier and provide less cooling when used as a single unit.
A surprising aspect is that AI data centers do this despite the uncertainty surrounding the AI chip supply of the United States. The White House recently NVIDIA H20, in accordance with China, And AMD Mi308 chips For export to China, and there are also rumors that the 5090D could also be affected by the ban. In fact, the rumor says that Nvidia mentioned the suspension of providing chips 5090D to its partners of the board of directors, although it did not mention the entirely sales stop.
Despite this, companies always abandon the excess capacity – probably because they have to cover their financial costs, or they may pass. In addition, as flea technologies are advancing, companies that use older generation GPUs will no longer be competitive, and they will be forced to sell these cards anyway. Therefore, it is probably logical that they disclose these underused assets now, then buy everything available when the request really occurs.