Chinese OEM Unveils GeForce RTX 5090 with 96GB VRAM at a Fraction of the Cost

Shenzhen Suqiao Intelligent Technology, a well-established OEM and ODM specializing in graphics cards, motherboards, and servers, has introduced a modified GeForce RTX 5090 featuring an impressive 96GB of memory. This is three times the 32GB found on the standard retail version. The card is currently listed on Alibaba for $3,888, which is approximately 35% less than the typical retail price of an RTX 5090 in the United States. For comparison, NVIDIA’s professional-grade RTX PRO 6000 Blackwell, which utilizes the same GB202 GPU die, starts at a staggering $16,000.

Custom Engineering and Memory Expansion

Suqiao appears to have achieved this expanded memory capacity by designing a custom PCB that employs clamshell memory architecture. This technique involves mounting GDDR7 modules on both sides of the board, effectively doubling the available VRAM—a method similar to what NVIDIA uses in its Pro 6000 series. The GPU itself may be sourced from harvested RTX 5090 units, which are then reballed onto the new PCB and operated using modified or leaked firmware. This approach has precedent, as similar modifications were previously used by Chinese workshops to upgrade RTX 4090 cards to 48GB of VRAM.

However, there are some inconsistencies in the product listing. Notably, it references GDDR6X memory running at 14 Gbps, which does not align with the specifications of a genuine 96GB Blackwell-based card. This discrepancy raises questions about the card’s authenticity and actual performance capabilities.

Market Impact and Considerations for AI Developers

Other vendors on Alibaba are reportedly offering similar modified RTX 5090 cards for around $5,900, positioning Suqiao’s offering at the lower end of this emerging gray market. If these cards are legitimate, they could provide AI researchers and developers with access to large VRAM pools at a significantly reduced cost compared to official workstation-class GPUs. However, potential buyers should be aware of the trade-offs, including possible firmware instability and the absence of official support or warranty.

As the demand for high-memory GPUs continues to grow, especially in AI and data science applications, these unconventional solutions highlight both the innovation and risks present in the current graphics card market.