A number of major semiconductor companies, such as NVIDIA, AMD, and Intel, have been severely harmed by the new U.S. government limits on AI hardware exports to China, which prevent them from offering high-performance AI processors for sale in China. Because of their powerful computing power, NVIDIA’s GeForce RTX 4090 gaming GPUs have been significantly impacted by this ban and are now unavailable in mainland China. As we previously reported, NVIDIA allegedly relocated a sizable stock of its AD102 GPUs and GeForce RTX 4090 graphics cards to China ahead of these restrictions. This might have added to the worldwide scarcity of RTX 4090, pushing up the cost of these cards to USD 2000. Unexpectedly, sources from the Chinese Baidu forums have revealed that specialized factories all throughout China are turning these GPUs—which came in before the ban—into artificial intelligence (AI) solutions.

This conversion entails taking apart the game GPUs, taking off the cooling mechanisms, and removing the GDDR6X memory and AD102 GPU from the main PCBs. Then, these parts are re-soldered onto a “reference” PCB made in the United States that is better suited for AI applications and has dual-slot blower-style cooling made for server environments. The blower-style cooler is only two slots wide, and many of them can be put in parallel in an AI server, while the third-party coolers that come with these GPUs are three to four slots in size. These redesigned RTX 4090 AI solutions are given to Chinese businesses using AI workloads following extensive testing. Because the primary GPU and memory components have been removed, this adaptation process has led to an inflow of RTX 4090 coolers and bare PCBs at remarkably low prices into the Chinese reseller market.

The disassembly of AIB GPUs before they are transformed into blower-style AI server-friendly graphics cards is seen below.

Factories are stress-testing these cards following modifications in Furmark, 3DMark, and AI applications to ensure that they function properly. They are also likely running some smaller LLMs locally to test the cards’ compute capability. Yes, if this process is not handled carefully, it could brick a lot of AD102 GPUs, but nobody outside these factories knows the yields of this experiment.