Chinese Firm Offers 96GB RTX 5090 Mod for $3,888
A Chinese manufacturer is selling a modified Nvidia graphics card with triple the standard memory at a price far below US retail. While the deal looks attractive for AI users, the listing contains technical inconsistencies that cast doubt on its authenticity.
Shenzhen Suqiao Intelligent Technology, an established Chinese hardware maker, has listed a modified Nvidia GeForce RTX 5090 on Alibaba for $3,888. The card reportedly features 96GB of video memory, which is three times the 32GB found in the standard US version. This price is roughly 65% of what consumers typically pay for the original flagship model in the United States, where custom versions start at $6,000.
The motivation behind such modifications is largely driven by the boom in artificial intelligence. Many users need more memory to run large language models locally, and standard graphics cards often fall short. Chinese factories have a history of upgrading older cards with more memory, such as the RTX 4090 and 3090, to meet this demand. While this specific 96GB claim seems technically possible given the shared silicon with Nvidia’s professional RTX Pro 6000, the details provided in the listing raise significant red flags regarding the actual components used.
Technical specs contradict known memory standards
Tom's Hardware points out a critical inconsistency in the product description. The listing claims the card uses GDDR6X memory running at 14 Gbps. However, GDDR6X chips are typically much faster, operating between 19 and 24 Gbps. Furthermore, the largest GDDR6X chips produced by manufacturers like Micron are 2GB each. Even if the card used the maximum possible number of chips on both sides of the circuit board, the total capacity would only reach 48GB, not the claimed 96GB.
To actually achieve 96GB, the card would need to use GDDR7 memory, which is faster and denser. The RTX Pro 6000 Blackwell uses GDDR7 to hit that 96GB mark. If Suqiao is indeed using GDDR6X, the 96GB claim is physically impossible with current chip capacities. This discrepancy suggests the listing may be misleading, or the seller is confusing memory types, which is a common issue in gray-market hardware sales.
Firmware hacks enable higher memory limits
Getting a graphics card to recognize more memory than it was designed for requires more than just soldering new chips. It necessitates modifying the card’s firmware. In the past, modders have successfully increased the memory of cards like the RTX 4090 by hacking the firmware. Rumors suggest that leaked firmware for the RTX 5090 has been circulating for months, which could explain why these modified units are starting to appear on the market.
However, buying such a card carries risks. The seller is a legitimate OEM, but the specifications listed are questionable. If the memory is actually slower GDDR6 or GDDR6X rather than the faster GDDR7, the performance in AI workloads might be significantly lower than expected. Buyers would be paying for capacity that does not match the speed required for modern inference tasks.
Price advantage comes with reliability risks
The $3,888 price point is undeniably attractive compared to the $6,000+ cost of official custom RTX 5090s in the US. Even if the price were higher, at $5,900 as listed by other merchants, it would still offer a substantial saving for those needing high memory capacity. The trade-off, however, is the lack of official support and the potential for hardware instability due to unofficial modifications.
For professionals relying on these cards for critical AI workloads, the risk of driver incompatibility or hardware failure may outweigh the cost savings. The situation highlights the gray area of the GPU market, where demand for AI resources drives manufacturers to push the limits of existing hardware. Until the technical details are verified, the 96GB claim remains suspect, and buyers should proceed with extreme caution.
