Upgraded Nvidia RTX 5090 gets 128GB VRAM and $13,000 price tag — 'super limited' GPU is described as a 'prototype'
Allegedly powered by a custom PCB and prototype GDDR7 memory

We've seen our fair share of modified GPUs from China — a region restricted from receiving most high-end cards. These export controls imposed by the The White House have created an underground market where gaming-grade graphics cards are stripped of their internals and put onto custom PCBs designed for AI workloads. AI requires a lot of memory, and despite former flagship GPUs like the RTX 4090 already coming with large memory buffers, it's still not enough for tasks such as running LLMs locally.
The latest to join the line of modded GPUs is the RTX 5090 with a whopping 128 GB of VRAM. Yes, this graphics card most likely has more memory than your entire PC. It's described as a prototype because the modder is using GDDR7X memory that is not yet available on a larger scale, but there's still a "super limited" quantity associated with this, which suggests that more units could be made.
Another fun 5090 128G prototype 😂 super limited with ~13.2k USD price 😂 pic.twitter.com/d0ypVglfIISeptember 7, 2025
By default, the RTX 5090 comes with 32GB of GDDR7 memory, which might seem plenty, but clearly there's room for improvement. What's interesting is this mod is a tier above anything we've seen before, or thought was possible. Previously, an RTX 4090 was upgraded with 96 GB of VRAM by using the best GDDR7 modules available in a dual-sided config. That was an anomaly, too, as the usual modded 4090s top out at about 48GB VRAM. The reason behind which is the complexity of the custom PCBs that are needed to support such tampering.
Currently, the highest-density GDDR7 chips are 24Gb (or 3 GB) per module, which can allow for up to 48 GB of VRAM. To reach higher capacities like 96 GB, dual-sided PCBs are used with memory modules on top and bottom, as seen on Nvidia’s RTX Pro 6000. Therefore, for an RTX 5090 to feature 128 GB of VRAM, the factory would either need access to unreleased 32 Gbit GDDR7 chips — which no vendor has announced yet — or design a Frankenstein custom PCB with far more memory chips than usual; 64 GB on either side, totaling 128 GB in a dual-sided config.
Not only that, but custom firmware and BIOS are also required for said VRAM to be properly recognized and utilized by your system. The screenshot in the source tweet is from Nvidia-SMI which clearly shows the full 128GB memory being detected, meaning the mod works and is not just proof-of-concept. Moreover, the fact that there's pricing for this card is even more evidence that it's a potentially commercially-viable product ready to be sold.
According to I_Leak_VN, the 128 GB 5090 is priced at roughly $13,200 which is more than 4x the street pricing of a regular 5090. Shockingly, it almost makes sense considering the 96 GB RTX 6000 Pro sells for $10,000. If you do the math, it's 33% more memory for 33% more price with the souped-up 5090. Also, none of that accounts for the actual "craftsmanship" of performing the mod by hand. Regardless of the illicit nature of these GPUs, someone has to work diligently to make them possible, and it's not easy to casually swap GPU cores and VRAM chips.
Unfortunately, after the core and memory is extracted from a normal 5090, they often make their way onto secondhand marketplaces where they're sold as real 5090s to scam buyers. It's important to stay vigilant, especially now that the AI boom has created this new precedent where various hollow 4090s and 5090s have flooded the market for cheap, so if anything looks to good to be true, it likely is.
Get Tom's Hardware's best news and in-depth reviews, straight to your inbox.
Follow Tom's Hardware on Google News, or add us as a preferred source, to get our up-to-date news, analysis, and reviews in your feeds. Make sure to click the Follow button!

Hassam Nasir is a die-hard hardware enthusiast with years of experience as a tech editor and writer, focusing on detailed CPU comparisons and general hardware news. When he’s not working, you’ll find him bending tubes for his ever-evolving custom water-loop gaming rig or benchmarking the latest CPUs and GPUs just for fun.
-
bit_user
No, the max number of chips is 2 per channel, for GDDR memory. Its timing is so tight that the way double-sided (i.e. "clamshell") configurations work is by placing each pair of chips exactly opposite each other, so that the signal timing is identical.The article said:Therefore, for an RTX 5090 to feature 128 GB of VRAM, the factory would either need access to unreleased 32 Gbit GDDR7 chips — which no vendor has announced yet — or design a Frankenstein custom PCB with far more memory chips than usual; 64 GB on either side, totaling 128 GB in a dual-sided config.
Therefore, the only way such a board should exist is if it uses 32 Gigabit packages (i.e. 4 GiB) and uses two per 32-bit channel (i.e. in clamshell mode). What we probably can't rule out is the possibility of die-stacking, within the GDDR7 packages. That might be how they reach 32 Gigabits per package, even with no 32 Gigabit dies available. If that's what they did, perhaps a 192 GB board isn't far behind.
Before anyone gets too enthusiastic about this, I'd expect memory frequencies to have been scaled back, in order to make it work. Consequently, the gaming experience on existing 32 GB cards is probably the best of the lot. -
usertests
Yeah this is very questionable. I think we'd know if 4 GiB packages were imminent. I think we've seen them on a roadmap or two, but it's too soon after 3 GiB GDDR7.bit_user said:Therefore, the only way such a board should exist is if it uses 32 Gigabit packages (i.e. 4 GiB) and uses two per 32-bit channel (i.e. in clamshell mode). What we probably can't rule out is the possibility of die-stacking, within the GDDR7 packages. That might be how they reach 32 Gigabits per package, even with no 32 Gigabit dies available. If that's what they did, perhaps a 192 GB board isn't far behind.
Is this die stacking trick even possible for a third party to do (i.e. not Samsung, Micron, SK Hynix)? -
bit_user
Don't know, but it was the only thing that came to mind that could enable a big jump vs. their roadmaps.usertests said:Is this die stacking trick even possible for a third party to do (i.e. not Samsung, Micron, SK Hynix)?
TBH, I'm really skeptical the chips are even setup for stacking. Perhaps more likely would be some sort of active interposer, with two GDDR7 dies lined up side-by-side. I'm really just grasping at straws, here. -
aberkae Someone has to tell AMD they are underhyping ai to compete with the 5090 and rtx 6000 pro! 😅Reply -
usertests
128 GB is old news now. We need a new Halo with 384 GB (and AI gaming handhelds with 384 GB of RAM).aberkae said:Someone has to tell AMD they are underhyping ai to compete with the 5090 and rtx 6000 pro! 😅 -
Makaveli
You going to pay 20k for that :)usertests said:128 GB is old news now. We need a new Halo with 384 GB (and AI gaming handhelds with 384 GB of RAM). -
abufrejoval I resent the usage of the term "illicit" in the article: there is nothing illegal about those cards.Reply -
bit_user
If it's not being done with Nvidia's blessing, then they aren't totally licit!abufrejoval said:I resent the usage of the term "illicit" in the article: there is nothing illegal about those cards.
The issue could be that Nvidia finds their supply and cuts it off. Or, maybe Nvidia puts something in their firmware (which is cryptographically signed) that refuses to run or support such a quantity of memory. Or at the driver level.
There are lots of ways Nvidia could cause problems for them, if it wanted to. -
JarredWaltonGPU
I'm more inclined to believe this is a scam to dupe some unsuspecting buyer rather than a real GPU using "prototype" boards or memory or whatever. Even if it does work, that prototype nature could mean stability is severely compromised. I wouldn't touch this thing with a ten foot pole, in other words!bit_user said:No, the max number of chips is 2 per channel, for GDDR memory. Its timing is so tight that the way double-sided (i.e. "clamshell") configurations work is by placing each pair of chips exactly opposite each other, so that the signal timing is identical.
Therefore, the only way such a board should exist is if it uses 32 Gigabit packages (i.e. 4 GiB) and uses two per 32-bit channel (i.e. in clamshell mode). What we probably can't rule out is the possibility of die-stacking, within the GDDR7 packages. That might be how they reach 32 Gigabits per package, even with no 32 Gigabit dies available. If that's what they did, perhaps a 192 GB board isn't far behind.
Before anyone gets too enthusiastic about this, I'd expect memory frequencies to have been scaled back, in order to make it work. Consequently, the gaming experience on existing 32 GB cards is probably the best of the lot.