Chinese factory reportedly dismantles RTX 5090 in bulk to convert them into AI accelerator cards

The latest leaked video was shared by Hong Kong hardware media outlet HKEPC. The footage shows a large-scale modification workshop piled with large quantities of retail RTX 5090 graphics cards, where workers, following an assembly-line process, disassemble the cards, remove the original coolers, and transfer the GPU and memory chips onto custom circuit boards. Graphics cards from manufacturers such as PNY, Gigabyte, and ASUS can be seen in the video, including even Nvidia's own Founders Edition reference products.

Judging from the on-site footage, this is not an individual modification performed by an ordinary computer enthusiast on a workbench, but a production operation of a certain scale. Large numbers of graphics cards are processed in a centralized manner, and the modified products are also tested to confirm that they can operate normally. The original video is said to have been posted on social media and subsequently deleted, but HKEPC retained the relevant footage.

The core reason for the modification is that the design of ordinary gaming graphics cards differs markedly from the needs of data centers. The RTX 5090 uses a bulky multi-fan cooler that typically occupies three or more expansion slots. In an ordinary desktop, this design is conducive to heat dissipation, but in server chassis that require densely installing multiple GPUs, it takes up too much space and makes organizing airflow inside the chassis more difficult.

Therefore, modification workshops remove the original GPU cores and memory chips from retail circuit boards and solder them onto redesigned custom PCBs. The new circuit boards usually adopt a more compact dual-slot design and are equipped with blower-style coolers that direct hot air out of the chassis in a concentrated manner. In this way, graphics cards originally designed for gaming PCs become better suited for installation in multi-GPU servers or AI computing systems.

This type of modification is not simply replacing the cooler. GPU cores and memory chips use complex packaging and soldering processes, requiring professional equipment for removal, cleaning, reballing, and soldering. Custom circuit boards must also redesign the power delivery traces, component layout, cooling structure, and interface positions to meet the operating requirements of high-power GPUs. After modification is completed, the products also need to undergo functional testing and stability verification before they can be used for actual computing tasks.

The RTX 5090 uses Nvidia's Blackwell architecture, is equipped with a GB202 GPU core, and features 32GB of GDDR7 memory. Its memory capacity and computing capability enable it to handle some large-model inference, model fine-tuning, and other AI computing tasks. For enterprises or research teams unable to obtain certain data-center GPUs through official channels, converting consumer-grade GPUs into server accelerator cards may become an alternative way to obtain computing resources.

However, RTX 5090 modified cards cannot be fully equated with data-center GPUs specially designed by Nvidia. The two may differ in memory capacity, cooling redundancy, reliability design, server management functions, software support, and long-term operation guarantees. The modified RTX 5090 still mainly retains the original 32GB of GDDR7 memory, and will not automatically gain larger capacity or all the functions of data-center products simply because a custom circuit board is installed.

Some recent claims say that Chinese manufacturers can expand the RTX 5090's memory to 64GB or even 96GB, but there is currently no reliable public evidence proving that these high-capacity modified versions have achieved stable operation. Judging from existing information, the products shown in the latest video still use the original 32GB memory configuration.

This phenomenon is also related to U.S. export restrictions on advanced computing chips. U.S. export controls on China have restricted certain high-performance computing chips from entering the Chinese market, while the computing performance of the consumer-grade RTX 5090 is highly attractive. As a result, related products are converted into accelerator cards suitable for AI servers through disassembly and circuit board redesign, sparking discussion about the actual effectiveness of export controls.

However, based solely on the appearance and quantity of the products in the video, it is still impossible to determine the specific procurement channels for these graphics cards, whether the transactions are legal, or whether they violate relevant export control regulations. The modification itself, the source of the chips, the cross-border transportation methods, and the final use are separate issues that need to be judged based on specific facts.

Meanwhile, the supply situation of the RTX 5090 in the global market has also attracted attention. Recent reports say that this flagship gaming graphics card has become extremely difficult to buy in some markets, with retail prices significantly higher than the official suggested price. Nvidia executive Jeff Fisher said in mid-October that the RTX 5090 has not been discontinued and the company will continue to supply it, but supply chain capacity is indeed constrained. Therefore, the current tight supply of graphics cards cannot simply be attributed to all products having stopped production.

If large numbers of RTX 5090s are used for AI computing rather than entering the ordinary gaming PC market, this may further worsen the difficulty players face in buying flagship graphics cards. However, the video does not provide the factory's actual output, procurement scale, or market share, so its impact on global retail supply cannot yet be quantified.

For Nvidia, this kind of modification also reflects the increasingly blurred boundary between consumer-grade GPUs and AI accelerators. Although the RTX 5090 is aimed at the gaming market, its powerful parallel computing capability is also suitable for some AI workloads. As long as the GPU core and memory can operate stably, redesigning the circuit board and cooling system gives it a chance to enter server environments for which it was not originally designed.

Overall, the latest video shows that Chinese manufacturers are converting retail RTX 5090s into accelerator cards more suitable for AI servers in a relatively industrialized manner. This not only reflects market demand for high-performance GPU computing capability, but also shows that consumer-grade hardware can enter the professional computing field through redesign. However, there is still insufficient public data on the actual scale of the modified products, the source of the chips, long-term reliability, and their impact on the supply of gaming graphics cards.