Chinese Firms Convert RTX 5090 GPUs into AI Accelerators, Bypassing Export Restrictions

RTX 5090s get ready to blow with AI-focused makeovers in China -- industrial production lines transplant GPU and memory from gaming cards onto server-ready PCBs with blower-style coolers

The RTX 5090 is supposed to be the crown jewel of gaming GPUs -- a 600W monster designed for enthusiasts. But in China, a very different story is unfolding. A recent video from Bilibili creator Ajian Talk exposes a parallel production line where these consumer cards are stripped bare, rebuilt, and reborn as blower-style "Turbo" models for AI servers. We've seen AI-focused 5090s surfacing from China before, but this is a rare look into the actual process of how they're made. What's striking is the scale at which this work is done, as Ajian's footage shows stacks of retail RTX 5090 boxes, complete with packaging and protective films, sitting in storage. These cards are tightly controlled and only available as 'D' variants due to U.S. export restrictions, yet somehow a Shenzhen-based company called "CT" has stockpiles large enough to fuel entire assembly lines. These are standard gaming variants from the likes of Zotac, Palit, and Inno3D, too. The process starts with full retail 5090s undergoing functional tests; workers check the GPU die and memory to ensure they're stable. Any defective boards are discarded before a single screw is turned. For the rest, CT tears off the triple-fan coolers and preps the "full-blood" GB202 chips for transplant. In the video, the host says that contrary to online rumors, blower cards aren't made with downgraded silicon -- only the best binned, high-performing dies are selected for this conversion. There are some "bad" dies on gaming 5090s that are completely discarded. Once the testing and teardown process is complete, it's time to do the actual work. The conversion is mostly straightforward, but it's not just manual tinkering. Ajian's video reveals industrial-grade BGA rework stations lifting the GPU dies and VRAM modules, cleaning them, and reballing them for mounting on custom dual-slot PCBs designed specifically for blower cooling. Every component -- from capacitors to power connectors -- is placed with 15-micron precision by SMT machines, while nitrogen-fed reflow ovens solder everything under controlled conditions to avoid oxidation. After reassembly, the cards are fitted with blower coolers, metal backplates, and server-grade I/O shields. The RTX 5090 Turbo is born. Each card undergoes 24-hour burn-in tests and repeated power cycles, ensuring it can survive the brutal demands of AI workloads. GPUs with blower-style coolers work well when stacked densely, as they help the hot air escape out the back instead of circulating it, like in a normal case with an axial cooler. Blower GPUs like the 5090 Turbo are thinner (just 2 slots) and designed for consistent airflow in rack-mounted systems. Up to eight cards can fit into a single chassis, making them ideal for AI training clusters. CT's models reportedly use higher-grade components, better PCB design, and improved gold plating on PCIe contacts, all to guarantee stability at 600W power levels. The video also highlights a deeper issue: consumer GPUs are being funneled into AI markets, echoing the shortages seen during the crypto mining boom. For gamers already struggling to buy a 5090 at retail prices, these conversions might feel like history repeating itself. Still, the efficiency of CT's operation is remarkable and a glimpse at how quickly consumer hardware can be repurposed for enterprise needs. Whether NVIDIA condones this or turns a blind eye is another story, we've not seen the company pursue any investigation into these illicit operations before. In the end, the RTX 5090 Turbo becomes more than a simple shroud swap -- it's a reflection of how the GPU market is shifting under the weight of AI demand. Originally looked at as a symbol of gaming excess, the 5090 has now been reshaped for data centers, blurring the boundary between consumer tech and industrial hardware.

Chinese video shows factory-scale conversion of RTX 5090 cards into custom AI accelerators

Export restrictions are bypassed as RTX 5090 boards are reworked into rack-ready AI modules A video circulating on Bilibili shows how enterprising Chinese technicians are converting GeForce RTX 5090 graphics cards into high-efficiency AI accelerators, bypassing export restrictions in the process. Instead of using the retail GPUs as is, a local manufacturer is removing the GB202 chips and memory modules from the gaming boards. The components are then transferred onto custom two-slot PCBs and paired with blower-style cooling, optimized for server rack environments. The operation appears to be being run at factory-level scale. Robotic arms extract GPU modules, pick-and-place systems reassemble them onto redesigned boards, and automated soldering equipment finalizes the electrical work, before visual inspection systems confirm alignment and build quality. The process is fast, repeatable, and efficient, far beyond what might be expected of a small workshop, but what makes this enterprise even more notable is the sheer volume of RTX 5090 cards seen in the video, despite restrictions on their export to China. Nvidia has not officially released the 5090D variant, designed to comply with U.S. sanctions, and yet shelves of full-power RTX 5090 units can be seen undergoing retrofitting. Gaming models with three-fan cooling are too large for server chassis, and their power connectors are often positioned awkwardly for rack use, so the cards are physically modified to fit data center constraints. The redesigned blower cards shift the power input to the tail end of the PCB, simplifying cable routing and avoiding risk of overheating. From a hardware standpoint, the GPUs function identically. By shifting them onto slimline, server-compatible boards, Chinese firms are essentially creating their own high-end AI accelerators. If hardware intended for gamers can be so easily adapted for AI research and model training, then questions about the effectiveness of the current export controls surely have to be asked. It's clear from the scale of the operation in the video that there is strong demand for AI accelerators built from repurposed RTX 5090 cards. For AI developers in China, it means they can get the computing power they need without having to rely on official supply channels.

China Has Automated Conversion of RTX 5090 Cards into AI Accelerators

A video circulating on the Chinese tech community's Bilibili platform offers an inside look at how a local manufacturer is converting NVIDIA's RTX 5090 gaming GPUs into dedicated AI accelerators. Instead of leaving the factory‑shipped circuit boards intact, workers use automated equipment to extract the GB202 GPU modules along with their onboard memory. These modules are then placed onto custom dual‑slot PCBs, which are paired with blower‑style heatsinks -- an ideal setup for cramped server racks and dense AI clusters. What's surprising isn't just the level of automation, but also the sheer volume of cards. The footage shows stacks upon stacks of retail RTX 5090 units, even though U.S. sanctions prohibit the official distribution of this model in China. NVIDIA did design an RTX 5090D variant to meet compliance requirements, yet that version hasn't been made available in the domestic market either. It appears that unofficial supply routes are channeling unmodified RTX 5090 boards into local factories, where they undergo this retrofit process. The technical workflow is notable for its efficiency. Robotic arms handle the precise removal of GPU and memory chips, after which pick‑and‑place machines position them onto the new board. Automated soldering stations complete the electrical connections, and finally, quality‑control cameras perform visual inspections. By minimizing manual intervention, the manufacturer can maintain high throughput and consistent build quality -- characteristics more akin to a formal GPU production line than a small‑scale workshop. This phenomenon highlights a key challenge in controlling the spread of advanced AI hardware. Even when top‑tier GPUs are officially restricted, demand from AI researchers and data centers can drive alternative markets. The operation showcased in the video isn't a one‑off project; its scale and sophistication imply an organized network that prioritizes custom‑built AI solutions. In effect, the process undercuts the intended impact of export controls by enabling continuous access to cutting‑edge hardware. For China's AI ecosystem, having a steady supply of powerful GPUs is crucial. Training large neural networks demands vast computational resources, often only available through high‑end accelerators like the RTX 5090. By repurposing gaming cards into server‑grade modules, local companies can sidestep official channels while still meeting performance requirements. This workaround could accelerate AI research but also creates a regulatory blind spot. Ultimately, the video demonstrates that even rigorous export restrictions can be challenged through technical ingenuity and scale. It raises important questions about how global policymakers can adapt to such workarounds, especially as AI continues to grow as a strategic priority worldwide.

A Huge Inventory Of GeForce RTX 5090 Is Getting Turned Into AI GPUs In China; GPU And VRAM Chips Removed And Installed On New Boards With Blower-Style Coolers

A mass conversion of GeForce RTX 5090 cards is happening in China through machines, which are turning the "Gaming" cards into "AI"-specific ones. Despite tight regulations and strict export policies, some companies seem to have no problem in getting the GeForce RTX 5090 in China. While the GeForce RTX 5090 D is getting a replacement with the RTX 5090 D V2, the RTX 5090 is somehow smuggled to China, where they don't actually end up in someone's gaming rig but in AI systems after going through a big modification. This is what we saw in a video, where a Chinese company is transforming all the GeForce RTX 5090 GPUs into AI cards through automated processes. The lab where all this stuff is going on is doing this on a mass scale and has probably transformed thousands of RTX 5090s into dedicated AI cards. First, the GPUs are tested to see if they are fully functional, and then their GPU chip (GB202 die) and the GDDR7 VRAM modules are removed from the PCB. They solder the GPU chip and VRAM chips on a custom PCB made specially for making the new AI cards. The PCB is quite large and even has a 16-pin power connector placed at the right edge for an easier connection in AI systems. These are then coupled with large AI blower-style coolers, which enhance the cooling of the components so that they can operate for longer sessions. Almost everything is automated, from removing the chips to installing them on the new PCBs, which are done through machines with precision. They are then tested again and packed for shipping to firms. This is how a lot of RTX 5090s are converted to AI cards, as, unlike the RTX 5090 D, the RTX 5090 retains its AI capabilities, which are quite popular among enterprises. Since the upcoming RTX 5090 D V2 won't likely have such AI features due to US restrictions and will also have a nerfed memory configuration, it will likely end up in gamers' rigs. However, the flow of the RTX 5090 into the Chinese market doesn't seem to have a problem at all, and if you remember, this isn't the first time we have seen such high-end GPUs getting converted into AI cards. In 2023, we also saw NVIDIA's flagship GeForce RTX 4090 GPUs turning into AI cards with similar blower card solutions. However, the RTX 4090 was initially allowed to be sold in China, but the mass conversion seems to have happened later.

GeForce RTX 5090 Reportedly Has Easy Availability In China; After Conversion To Blower Style, No GPUs Suffers From Connector Melting

The Chinese market apparently has no problem with a steady supply of the GeForce RTX 5090, which is banned due to US export policies. Just a few days ago, we reported on the NVIDIA GeForce RTX 5090 smuggled to China, which is officially banned due to US export policies. The GPUs in the lab get turned into dual-slot blower-style cards, which are then deployed in the servers for AI workloads. The Bilibili channel that posted the video has provided more information about this whole process in the comments, revealing that the GeForce RTX 5090 is easier to get in China. As per the report, the availability of the card is stable and is also available on major e-commerce platforms. This means a lot of RTX 5090 inventory is readily available to the enterprises, since the RTX 5090 offers great AI performance, unlike the RTX 5090 D, which was nerfed due to US policies. Nonetheless, the new RTX 5090 D V2 is coming next week, but enterprises won't need that due to the stable availability of the faster RTX 5090, which has superior AI capabilities and 32 GB VRAM compared to 24 GB on the RTX 5090 D V2. According to the Bilibili channel 阿健座谈, which posted the video, the RTX 5090 in its custom design isn't ideal for server deployment. The server cases reportedly need something more compact, and since they receive the custom designs, which are often 2.5-3.0 slots thicker, they are turned into dual-slot GPUs with a blower-style cooler, but that's done after installing the GPU chip and VRAM modules on a newer PCB that has the 16-pin connector on the right. Once again, this is done since the server cases don't have that 35mm clearance for the 16-pin connector, which is crucial for preventing any melting issues. When the connector is positioned on the right, not only does it save space, but the melting problem is almost eliminated. As per an enterprise, not a single RTX 5090 saw the connector melting so far, even after deploying nearly a thousand RTX 5090 and RTX 4090s. This also goes in line with what EVGA's Kingpin thought about the top location of the connector on these high-end GPUs. Not only do these require more clearance for protecting the cable from a bend at the end, but it's also closer to the VRM and runs hotter than being on the right edge of the PCB.