Some AI data centers in China are reportedly holding large stocks of China-specific 48GB Nvidia RTX 4090D GPUs, dismantling and refurbishing them, and then reselling them on the market as new cards. DigiTimes Asia reports that a few companies are resorting to letting go of excess computing capacity to generate multiple times the profit compared to renting the GPUs out, which will take them about three to five years to recoup their investment.
This move is a sign of how China’s AI rush is leading to billions of dollars in idle infrastructure. The report states that an AI data center requires a utilization rate of more than 70% to 75% for it to turn a profit. However, activation rates remain below 20%, meaning a significant amount of capacity is left unused and many GPUs remain idle. To help them stay afloat, a few companies are turning to selling their unused assets to generate some quick cash and pay off the bank loans they used to purchase their hardware.
Selling RTX 4090 cards is quite lucrative, too. Currently, these China-specific RTX 4090D GPUs with 48GB of VRAM are priced between CNY20,000 and CNY40,000, or approximately US$2,735 and US$5,470. We’re unsure if these cards were used at all, but they still need to be modified if the data center wants to sell them to consumers. Data centers typically convert fan-cooled GPUs into blower-style cards, such as this blower-style RTX 5090D leaked on 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.
One surprising aspect is that AI data centers are doing this despite the uncertainty surrounding AI chip supply from the U.S. The White House has recently blocked China-compliant Nvidia H20 and AMD MI308 chips for export to China, and there are also some rumors that the 5090D might also be affected by the ban. In fact, it's rumored that Nvidia has mentioned suspending supplying 5090D chips to its board partners, although it did not mention stopping sales entirely.
Despite this, companies are still letting go of excess capacity — likely because they need to cover their financial costs, or they risk going under. Furthermore, as chip technologies advance, companies that use older-generation AI GPUs will no longer be competitive, and they will be forced to sell these cards anyway. Therefore, it probably makes sense for them to release these underutilized assets now and then purchase whatever is available when demand actually arises.
