We spotted a listing for a modified Nvidia GeForce RTX 5090 with 96 GB VRAM on China's Xianyu marketplace. It's important because it could give users a way to get massive amounts of memory at a much cheaper price than you'd pay for official pro cards. This could be really appealing for people looking to run large AI models on their own machines, even if it's a pretty risky technical proposition.
Unverified modification kit targets business AI clients with tripled memory capacity
The listing shows that someone's selling a full mod kit, including custom PCB schematics and modified BIOS firmware. The modification features a clamshell memory setup, where the memory (Samsung memory) is installed on both sides of the circuit board. That's designed to give you double the memory over the standard 32 GB you get with the retail RTX 5090.
Modification Specifications
- VRAM: 96 GB
- Memory Configuration: Clamshell configuration with Samsung memory modules
- Modification Package: Includes custom PCB schematics and modified BIOS firmware
They're asking 6,488,888 YUAN ($970,548) CNY ($970,548) – but that's just a starting point for negotiation. The seller makes it clear that they'll negotiate the price privately. And they're only selling this to business clients. The high price point seems to indicate this mod is aimed at businesses, not just hobbyists.
A 96 GB RTX 5090 could handle way bigger AI models than the standard 32 GB card. The mod looks like it's trying to give you the same AI acceleration power as the RTX Pro 6000 Blackwell, but at a lower price. But we haven't seen any independent confirmation that this specific 96 GB mod works as advertised yet.
There was another listing on Alibaba in September, but it had some conflicting memory specs that made us question its legitimacy. We've been following the GeForce RTX 5090 closely. And we've previously written about rumored Nvidia RTX 5080 24 GB. This current listing is still unconfirmed, so buyer beware until we get some independent testing to verify that this stuff really works.
Source: NotebookCheck
Discussion
0 comments