Modders double RTX 5080 memory to 32GB for AI workloads, raising supply concerns for gamers

Modders are slapping 32GB of VRAM on Nvidia's RTX 5080 GPUs, but that isn't good for gamers -- modded variants designed for AI workstations and servers

We have seen these types of mods on multiple generations of Nvidia cards; it was only inevitable that the RTX 5080 would get the same treatment. A report has surfaced overseas confirming that modders have figured out a way to double the RTX 5080's VRAM capacity from 16GB of GDDR7 memory to 32GB. This allows gaming GPUs to be used in workstations and servers, which could ultimately increase their use for AI workloads, potentially creating supply issues. Discovered by Uniko's Hardware (h/t VideoCardz), a Chinese GPU repair technician claims modders are "successfully" modifying RTX 5080s with 32GB of memory, and deploying these mods specifically on RTX 5080 cards boasting blower-style cooler designs that are optimized for workstations and servers. The modder did not explain how the modification process is being done. Likely, modders are adding eight additional 2GB memory chips to the RTX 5080's PCB to achieve 32GB of capacity. This is something Nvidia does on existing models, such as the RTX Pro 6000 and RTX 5060 Ti 16GB, to reach their respective memory capacities using a clamshell approach. This mod will significantly improve the RTX 5080's capabilities in AI-focused workloads. Depending on how many GPUs a system can house, a workstation equipped with up to four 32GB RTX 5080s can have up to 128GB of VRAM. AI model quality and performance depend heavily on memory capacity, so having more memory enables higher-quality and more complex models to be run. If this mod becomes popular, RTX 5080 supply could dwindle quickly as AI-minded buyers sweep up existing inventory, as happened with RTX 5090 sales. With the 32GB mod applied, the RTX 5080 can technically run the same AI models as the 32GB RTX 5090 (albeit with longer processing times), at just half to a third of the price (based on the constantly fluctuating prices of RTX 5090 SKUs). However, the popularity of this mod could be hamstrung by the availability of GDDR7 video memory ICs. The outgoing DRAM memory shortage has created exponential price increases and availability shortages across the entire industry. In 2025 alone, memory prices increased by 246%, and they're expected to keep rising well into 2026. Graphics cards have thankfully been hit the least among the major products that use memory, but ripple effects from the aforementioned availability issues are still affecting GPUs. For instance, in Japan, shops are rationing GPUs with 16GB or more of video memory. RTX 5080s modded with 32GB of VRAM will likely be made mainly from GDDR7 chips salvaged from damaged RTX 50-series graphics cards that repair technicians have on site. Regardless, it's cool to see this mod on the RTX 5080, even if it will inevitably be used almost entirely by AI-focused customers. That said, we will likely see some gamers with deep pockets have their RTX 5080s upgraded to 32GB of GDDR7, especially with the rumored RTX 50 Super refresh likely delayed by the ongoing DRAM shortage. If true, this 32GB mod will be the only way to upgrade the RTX 5080's memory until the RTX 5080 Super arrives with its rumored 24GB of GDDR7.

Modders are upgrading Nvidia RTX 5080 cards to 32GB of VRAM

Serving tech enthusiasts for over 25 years. TechSpot means tech analysis and advice you can trust. Hardware Slop?: While repairing a nonworking piece of gaming hardware is a common occurrence, adding extra RAM during the process is still unusual. However, a Chinese modder recently said that this kind of modification has now become widespread among enthusiasts. In China, specialized modders can add significantly more AI acceleration capabilities while repairing broken Blackwell GPUs. A technician claimed that turning standard GeForce RTX 5080 cards into 32GB monsters is already "a thing" in the country. If confirmed, the news could have some interesting consequences for the high-end GPU market. The GeForce RTX 5080 was originally introduced in January 2025, sporting a standard VRAM size of 16GB. So far, Nvidia is only providing 32GB with the biggest and baddest Blackwell model (RTX 5090), though the company is rumored to introduce a GeForce RTX 5080 Super model with 24GB of VRAM soon. According to more recent rumors, the alleged new GeForce model could arrive later than expected. The 24GB card may not arrive until the second quarter of the next year, following a soft launch that's internally scheduled for CES 2026. Even if these claims prove accurate, the new GPU will still have much less VRAM than the high-end GeForce RTX 5090 model. Chinese modders are allegedly working ahead of Nvidia's schedule, though. They are adding 16GB of VRAM to blower-type GPUs, which are designed to eject hot air from the rear. These types of cards are easier to turn into accelerators for AI inferencing and training workloads, expanding the overall computing capacity with multiple cards installed in the same custom server setup. The Chinese technician who disclosed the new modding/repairing trend said that the GeForce RTX 5080 32GB model is the most popular one. Other blower-style cards from the Blackwell generation (from RTX 5060Ti 16GB to RTX 5070Ti 16GB) are also popular models. The modders are likely exploiting information that emerged from Nvidia's 2022 hacking incident, using custom tools to make repair and upgrade work far easier. Judging from what I've read in my personal social circles, turning powerful gaming GPUs into yet another AI-focused hardware proposal is not exactly an exciting prospect for everyone. On Reddit, overclocking enthusiasts are highlighting the very sad state of the current RAM market, and the fact that turning even more hardware into "AI nonsense" is not a good modding goal to pursue. In any case, adding 16GB of VRAM to a 16GB GPU is still an interesting type of hardware surgery.

NVIDIA's RTX 5080 Is Now Being Sold in a "Chunky" 32GB VRAM Configuration in China, Gobbling Up GDDR7 Modules for AI Workloads

The Chinese GPU industry is reportedly selling modded GeForce RTX 5080s, which feature twice the memory compared to the GPUs' initial specifications. Given how desperate China's AI industry is to access compute power, one of the ways local AI setups have managed to fulfill their needs is by utilizing modified consumer GPUs, notably from NVIDIA. Since AI workloads require higher memory capacity onboard, Chinese modders have managed to adjust GPU configurations, and we have seen a similar trend with NVIDIA's RTX GeForce 4090. And now, according to @unikoshardware, NVIDIA's RTX 5080 GPUs are being sold with 32 GB GDDR7 memory onboard, in a workstation-oriented design. By the looks of it, such SKUs are highly popular among local buyers. Mounting the GeForce RTX 5080 with 32 GB of memory requires the integration of 3 GB GDDR7 memory modules, which we were expecting to see with the RTX 5080 SUPER, but Chinese modders are apparently one step ahead. The source mentions that the cards are being sold in a "turbo" style, which refers to the blower-fan designs seen with the modded RTX 4090s, as well. This allows for much more effective heat dissipation within these GPUs, which is a "must" considering how beefed-up modded units are with VRAM onboard. Such GPUs are widely used across local AI workloads, as Chinese customers have access to NVIDIA's consumer GPUs. By stripping them down and loading additional memory onboard, these GPUs are tailored for mid-end training/inference environments. As the modding community gains significant traction in China, Uniko's Hardware anticipates that such models could experience massive shortages once the domestic AI industry adopts GPUs with upgraded VRAM capacities. The more interesting aspect to observe here is how durable these modded configurations actually prove to be, given that power ratings are adjusted to be higher than factory-standard cards, which introduces a risk of its own. However, for the Chinese AI industry, it's probably more important to acquire computing capabilities than anything else, so durability might not be a concern for them.