China’s gray market ships 32GB-modified RTX 4080 SUPER cards for AI training—cheap capacity with risks around reliability, drivers, thermals, and support.
© A. Krivonosov
China has a new wave of unusual graphics cards: modified GeForce RTX 4080 SUPER units whose memory has been bumped from 16 GB to 32 GB using improvised methods. These boards aren’t affiliated with NVIDIA or any official vendor; third-party workshops rework them with a different audience in mind—not gamers, but startups and training farms for neural networks that need more memory while keeping costs down. The intent is straightforward: more capacity without premium pricing.
The cards have already appeared on Chinese marketplaces at around 9,200 yuan, roughly $1,200. Sellers offer them not one by one but by the box, a detail that hints at volume output. The shrouds carry a “CT AI” logo—the same company previously noticed moving large batches of modified RTX 3090 and 4090. Each unit is fitted with a new blower-style cooler, arrives wrapped in protective film, and even carries a sticker with a so-called “warranty” seal, though expecting real after-sales support seems unrealistic.
Taken together, this points to a growing gray market for AI accelerators in China: cheap, capable, and undeniably risky. These Frankensteins can be a sharp deal for compute-heavy workloads, yet questions linger over their reliability, driver compatibility, and thermal stability. The trade-off is obvious; the uncertainty is, too.