The RTX 5090 is vanishing from US shelves, and leaked footage suggests a good share of the cards are being pulled apart in Chinese workshops and rebuilt for AI servers.
The clips appeared in posts on X. As a hardware writer relaying them described the scenes, workers desolder the GB202 graphics chip and its GDDR7 memory from retail cards and move both onto a new board with a dual-slot blower cooler. Benches are stacked with boxed partner cards and Nvidia's own Founders Edition. Converted cards then go to test lines, where dozens run graphics and stability loads side by side before being repackaged for clients, likely data centers or AI labs. The same writer estimated that thousands of cards are converted every day. That is an inference from the footage, not a measured count.
The cooler swap is the point. Gaming flagships use thick triple-fan designs that occupy 3 to 4 slots and vent heat back into the case. That layout does not fit a 4U server packed with several cards. A blower pushes exhaust out the rear and fits in two slots.
Supply is tightening from the other side too. Nvidia has reportedly halted RTX 5090 production, with the writer guessing the silicon is going to higher-margin professional parts such as the RTX Pro 6000 Blackwell. Neither the halt nor the reason has been confirmed by the company in the material available. Meanwhile, custom RTX 5090s now start at $8,500 in the US, according to that writer, and the X posts put street prices near $10,000. Some prebuilt gaming PCs reportedly cost less than the bare card.
This pattern is familiar. Chinese shops converted RTX 4090s into blower-style cards after US limits on data center chips, because a consumer GPU is the cheaper route to similar compute. What differs now is the price gap. If a stripped card is worth more to an AI buyer than to a gamer, retail loses the bidding, whatever Nvidia intended. Nvidia's choice between serving gamers and chasing professional margins looks increasingly one-sided.