GeForce RTX 5090 Cards Stripped for AI Servers in Video From Hong Kong's HKEPC

In a workshop stacked with retail graphics card boxes, workers apply heat to boards and remove large GPU chips. Video shared on October 23, 2026 by HKEPC, a Hong Kong PC hardware outlet, shows NVIDIA's top-end graphics card, the GeForce RTX 5090, being taken apart one after another. The footage suggests that some of the scarce high-performance GPUs are being diverted not to gaming PCs but to AI servers.

At the workshop, workers removed the large coolers and heatsinks from retail cards and transplanted the GPU packages and GDDR7 memory chips onto a different printed circuit board (PCB). The finished cards were fitted with 2-slot blower-style coolers suited to dense server installation. In effect, it is a process for turning graphics cards designed for consumers into hardware for servers used in AI training and inference.

The video suggests this is not just an individual hobbyist's modification but work that may be happening at some scale. The likely backdrop is demand from AI workloads that need fast, high-capacity GPU memory. However, the footage alone cannot confirm the total number of cards modified, how they were sourced, or who ultimately receives them. The relationship to RTX 5090 shortages and U.S. export controls also requires separating confirmed facts from speculation.

HKEPC's Video of the RTX 5090 Modification Workshop

The video was shared by HKEPC, a Hong Kong-based hardware outlet. The footage it posted to X records a workshop lined with many GeForce RTX 5090 packages, along with the disassembly and modification process. Products from multiple manufacturers can be seen, including ASUS's ROG Astral series and NVIDIA's own Founders Edition. Workbenches and shelves hold large numbers of cards, suggesting a scale of work different from small individual modifications.

Workers first remove the retail card's large cooling assembly, then heat the board to remove the Blackwell-generation GB202 GPU package and the surrounding GDDR7 memory chips. Both the GPU and memory are mounted on the board using BGA (ball grid array) packaging, so removal requires dedicated heating equipment and precise handling. The recovered parts are then remounted on what appears to be a custom-designed PCB and paired with a 2-slot blower-style cooler. The finished cards are built to be used side by side inside a server chassis.

Still, all the video confirms is that a process exists for removing parts from retail cards and transplanting them onto another board. Even with many cards on screen, it is unclear whether this is a permanent production line handling tens of thousands of cards a year or a one-off batch of collected products. No independent verification data has been published on what share of modified cards work properly or whether they are actually being delivered to AI servers.

The regions where the cards were first sold, who bought them, and how they reached the workshop are also unknown. Some reports have suggested that RTX 5090 supply is being concentrated on AI servers, but this video alone cannot be used to judge the overall state of the market. What can be confirmed, at minimum, is that a route exists for taking high-performance GPUs sold as finished products, dismantling them, and reusing them as AI hardware.

Why the 32GB, 512-Bit Memory Configuration Drives the Modifications

Why would anyone take apart an expensive gaming graphics card and move it to another board? Answering that requires looking at the RTX 5090's hardware specifications and how they differ from the model made for the Chinese market. What matters most is GPU memory capacity and bandwidth, which determine how well AI models run.

According to NVIDIA's official specifications, the standard GeForce RTX 5090 has 21,760 CUDA cores and 32GB of GDDR7 memory. It has a 512-bit memory bus, a 28Gbps data rate, and bandwidth of 1,792GB/s, with a rated Total Graphics Power (TGP) of 575W for the whole card. In large language model (LLM) inference, performance depends not only on compute but on the memory capacity to hold model weights and the KV cache, and the speed at which they can be fed to the GPU. For workloads limited by memory bandwidth in particular, these specifications matter a great deal.

The GeForce RTX 5090 D v2 for the Chinese market, by contrast, keeps the CUDA core count at 21,760 but reduces the memory configuration. It has 24GB of capacity, a 384-bit bus, and 1,344GB/s of bandwidth, which is 8GB less capacity and 448GB/s less bandwidth than the standard model. NVIDIA's stated AI compute performance is also 2,375 AI TOPS for the D v2, versus 3,352 AI TOPS for the standard model. However, AI TOPS is a theoretical figure based on specific computing conditions and does not directly represent real-world LLM inference speed or throughput.

{
  "type": "bar",
  "title": "GeForce RTX 5090シリーズのメモリ帯域幅比較",
  "unit": "GB/s",
  "categories": ["RTX 5090（標準モデル）", "RTX 5090 D v2（中国向け）", "非公式ブロワー改造カード"],
  "series": [
    {"name": "メモリ帯域幅", "values": [1792, 1344, 1792]}
  ],
  "caption": "NVIDIAの公称仕様に基づく比較。改造カードの1,792GB/sは元の部品と同じ条件で動作した場合の理論値であり、実測値ではない",
  "source": "NVIDIA公式製品仕様"
}

Even with the same CUDA core count, the standard model and the D v2 differ clearly in memory capacity and bandwidth. With 32GB of memory, the standard model can keep more model weights and KV cache on the GPU, widening the options for quantization and for splitting models across multiple GPUs. Even for a 70B-class LLM, fitting the whole model into 32GB is difficult with typical 4-bit quantization, but it is possible to increase the portion of the model assigned to the GPU. That is also an advantage when combining multiple GPUs to run AI models.

Another reason is how well the cards fit into server chassis. Many retail RTX 5090s have large coolers occupying 3 to 4 slots, making it hard to line up several of them at tight spacing. NVIDIA's own Founders Edition is 2 slots thick, but it uses a "Double Flow Through" design that passes air through both sides of the card, a different cooling structure from a blower that concentrates exhaust at the back of the server.

Inside a server, it matters not only how thick the cards are but also that intake and exhaust directions are aligned. A 2-slot blower-style cooler makes it easier to keep spacing between adjacent cards narrow while pushing heat out the rear of the chassis. The main purpose of these modifications appears to be keeping the standard model's 32GB of memory while changing to a shape and exhaust structure suited to servers. Whether memory bandwidth and compute performance match the original card, however, depends on the modified board's design and operating settings.

Thermal Design and Reliability Problems of Modified Boards

Even if the GPU and memory can be transplanted to another board, that does not mean they will run stably in a server over the long term. Power delivery, cooling performance, and firmware compatibility are particular concerns. With retail cards, manufacturers design the board and cooling as an integrated whole and validate them for set operating conditions; with unofficial modified cards, it is unclear whether equivalent validation has been done.

The biggest worry is whether a thin cooler has enough capacity for the standard model's rated 575W power consumption. Retail RTX 5090s use large vapor chambers, heat pipes, and multiple fans to handle heat from the GPU, memory, and power circuitry. A 2-slot blower, by contrast, must expel a large amount of heat in a limited space. Combining high-speed fans with strong airflow from the server chassis might secure sufficient cooling, but how much power the cards can run at continuously is unknown.

Temperature management of the GPU package, GDDR7 memory, and power circuitry also becomes important in prolonged high-load operation. Insufficient cooling could reduce performance through thermal throttling as temperatures rise, and could increase thermal stress on components and solder joints. If the voltage regulator module (VRM) design differs from the original card, voltage stability and power delivery under heavy load could also be affected. Such issues have been discussed in past modifications of the RTX 3090 and RTX 4090, but no third-party data on temperature distribution, power control, or memory error rates has been found for these RTX 5090 boards.

Software and firmware compatibility is another challenge. A graphics card's operation depends not only on the GPU itself but also on the VBIOS (Video BIOS) on the board, power management circuitry, temperature sensors, and memory configuration. When the GPU and memory are moved to a different PCB, these configurations must be correctly recognized and work properly with NVIDIA's drivers. It is not known whether the modified cards use the original VBIOS unchanged or custom-modified firmware.

Furthermore, modifications that remove parts from a board are likely to fall outside the manufacturer's warranty. Replacement for failures and the supply of repair parts would fall to the modifiers or operators themselves. When deploying large numbers of GPUs in AI servers, not just whether they work but also failure rates, maintenance costs, and uptime affect operating costs. The video shows that the modification process exists, but it does not prove that the finished cards have reliability equal to enterprise products.

NVIDIA Denies RTX 5090 End of Life; Link to Shortages Unclear

The video drew attention against a backdrop of GeForce RTX 5090 shortages and rumors about its end of life (EOL). In October 2026, reports spread that production of the top-end GPU had ended and that GB202 chip supply would be concentrated on professional products. NVIDIA has denied this.

According to a report dated October 16 from Taiwan's Central News Agency (CNA), Jeff Fisher, NVIDIA's senior vice president for its personal computing business, said the rumor that the GeForce RTX 5090 is being discontinued is "not true." He called the RTX 5090 an important product and stated that shipments will continue. In an interview with 4Gamers, he also said that while the supply chain has capacity constraints, NVIDIA is working to keep supplying the product.

However, denying end of life is not the same as guaranteeing sufficient supply to the retail market. Fisher did not disclose specific future shipment volumes, regional allocations, or when the shortage would ease. The GeForce RTX 5090 was announced on January 6, 2025 and launched on January 30 at a suggested retail price of $1,999, yet poor availability was still being reported as of October 2026.

The video shows that some retail cards are flowing into modification routes for AI servers. But judging how much that affects the overall supply shortage would require comparing the quantities procured by modifiers against total RTX 5090 production and shipments. Such data has not been published, and it is not possible to conclude from the cards seen in the video that diversion to AI use is the main cause of the shortage.

GPU supply also involves multiple factors beyond semiconductor manufacturing capacity, including procurement of components such as memory and boards, production allocation among products, and retailers' inventory management. How much growing demand for AI products is affecting GeForce supply also cannot be quantified from public information. HKEPC's video is important as an example of high-performance consumer GPUs being repurposed for AI, but it is not decisive evidence explaining the global RTX 5090 shortage.

U.S. Export Controls and Opaque GPU Procurement Routes

This modification case also touches on a separate issue: U.S. export controls on advanced semiconductors. NVIDIA sells the GeForce RTX 5090 D v2, with lower memory capacity and bandwidth than the standard model, for the Chinese market. The cards in the video appear to be standard-specification RTX 5090s, and how they reached the modifiers has not been made clear.

On May 31, 2026, the U.S. Commerce Department's Bureau of Industry and Security (BIS) issued guidance on license requirements for advanced computing items. The document reaffirms that, under the U.S. Export Administration Regulations (EAR), exporting covered advanced computing items to companies headquartered in Country Group D:5 or Macau, or whose ultimate parent is headquartered there, generally requires a BIS license. Importantly, the license requirement applies based on the location of the ultimate parent even when the company is located in a third country.

According to BIS, this requirement was introduced on November 17, 2023 and continues to apply through subsequent rule changes. Covered items include advanced computing items classified under Export Control Classification Numbers (ECCN) such as 3A090 and 4A090. However, the May 31, 2026 guidance does not name the GeForce RTX 5090 directly. Whether a license is required for a given transaction depends on the product's export control classification, the destination, the end user, the use, and any applicable license exceptions.

{
  "type": "bar",
  "title": "標準RTX 5090と中国向けRTX 5090 D v2のメモリ帯域幅",
  "unit": "GB/s",
  "categories": ["RTX 5090（標準モデル）", "RTX 5090 D v2（中国向け）"],
  "series": [
    {"name": "公称メモリ帯域幅", "values": [1792, 1344]}
  ],
  "caption": "中国向けモデルは標準モデルよりメモリバス幅が128ビット、帯域幅が448GB/s小さい。輸出許可の要否は帯域幅だけでは判断できない",
  "source": "NVIDIA公式製品仕様"
}

If the RTX 5090 D v2, with its modified specifications, exists for China while standard models are reaching modifiers in China, the distribution route becomes an important subject of inquiry. However, the fact that a standard model was bought from a third-country distributor does not by itself establish a violation of export control rules. The legality of individual transactions cannot be judged without checking the exporter, consignee, end user, and whether required licenses were obtained. HKEPC's video does not allow such transaction details to be confirmed.

Some follow-up reports have also suggested that NVIDIA may be aware of and tolerating such modification workshops, or that it may be prioritizing supply to AI uses over the gaming market. However, no evidence has been presented that NVIDIA is involved in the unofficial modification process or profits from sales to modifiers beyond ordinary product sales. For now, the manufacturer's involvement must be distinguished from the fact that third parties are modifying products they bought on their own.

What this case highlights is the gap between the GPU performance AI uses demand and the products available through official supply channels. Where procurement of high-performance AI accelerators is constrained, there is an economic incentive to modify consumer GPUs with strong memory capacity and bandwidth and use them in servers. At the same time, once modification costs, failure risks, and export control constraints are considered, it is unclear whether this approach is viable in the long run.

Going forward, the key questions are how much power and performance the modified cards actually deliver, how durable they are, and through what transactions large numbers of RTX 5090s are being supplied to modifiers. Transplanting a GPU and memory onto another board shows one part of an unofficial effort to secure computing resources for AI. But clarifying its scale, market impact, and legal issues will require investigation based not just on video but on distribution records and hands-on testing of the actual cards.