Recently, a Shenzhen graphics card manufacturer called Suqiao Intelligent Technology (Suqiao) listed a modded Nvidia GeForce RTX 5090 on Alibaba, boosting its memory capacity from the standard 32GB all the way to 96GB, for only US$3,888 (about NT$126,000). That price is less than a quarter of the US$16,000 to US$17,999 price of Nvidia’s own same-capacity professional graphics card, the RTX Pro 6000 Blackwell. The emergence of this card once again highlights the maturity of China’s hardware-modding supply chain driven by demand for AI hardware, and exposes the huge price gap between Nvidia’s professional-grade and consumer-grade products. The standard RTX 5090 has a US MSRP of US$1,999 (about NT$65,000), but actual market prices have long been far above that figure.
Modding Principle: The Secret from Wafer to 96GB
The standard RTX 5090 features the GB202 chip and 32GB GDDR7 memory. To expand the memory to 96GB, Shuoqiao’s approach is to use a custom PCB design, soldering a total of 32 memory chips on both sides of the circuit board in “clamshell mode,” achieving three times the capacity of the original configuration.
As for the chip source, manufacturers may directly purchase GB202 silicon chips from Nvidia’s supply chain, or remove chips from existing RTX 5090 graphics cards, reball them, and solder them onto custom PCBs. This technique has precedent in China’s modding community; previously, the RTX 4090 was successfully modified into a 48GB version, and the RTX 3090 also has a 48GB modified version.
Spec concerns: GDDR6X or GDDR7?
However, there are obvious discrepancies in the product specification page on Alibaba. The page lists the memory type as GDDR6X and the speed as 14 Gbps, but these figures do not match the known specifications. The standard version of the RTX 5090 uses GDDR7, not GDDR6X; a speed of 14 Gbps corresponds to standard GDDR6, while GDDR6X actually operates at speeds between 19 and 24 Gbps.
In addition, the maximum capacity of the GDDR6X chips Micron produces is 16Gb (2GB), and even when configured in clamshell mode, they can only reach an upper limit of 48GB. To achieve 96GB, you would theoretically need to use 24Gb (3GB) GDDR7 chips, the same approach used in the RTX Pro 6000 Blackwell.
These conflicting specifications make it difficult for potential buyers to confirm the actual hardware configuration. As an OEM/ODM manufacturer with more than 10 years of experience in graphics cards, motherboards, and servers, Shuoqiao has a certain industry reputation, but the information on product specification pages should still be treated with caution. Before purchasing, it is advisable to ask the seller for actual hardware screenshots and test data.
Firmware Is Key: Leaked Boot Firmware
Nvidia graphics cards running non-standard memory configurations require modified firmware. The success of the modded RTX 4090 48GB version came down to modders modifying Nvidia’s firmware to support expanded memory addressing. The RTX 5090 96GB also requires similar firmware modifications, and there are even online reports of 128 GB.
Industry rumors indicate that leaked firmware for the RTX 5090 has been circulating in the modding community for months, which explains why modded versions are starting to appear on the market. The firmware must handle additional memory addressing and ensure compatibility with Nvidia drivers.
Price Comparison: How Tempting Is the Modified Version?
When you lay out the numbers, the price advantage of the modified version is very clear:
- RTX 5090 96GB Modded Version (Shuoqiao): US$3,888 (about NT$126,000; editor’s note: the relevant information can no longer be found, so it may just be a gimmick)
- RTX 5090 96GB Modified Version (Other Alibaba Sellers): About US$5,900 (about NT$192,000)
- RTX 5090 US Market Custom EditionStarting at US$6,000 (approx. NT$195,000)
- Nvidia RTX Pro 6000 Blackwell (96GB Official Version): US$16,000 to US$17,999 (approximately NT$520,000 to NT$585,000)
In other words, Shuoqiao’s modified version costs only about 24% of the RTX Pro 6000 Blackwell while offering the same 96GB of memory. For researchers and small teams that need a large memory pool to run local AI model inference, that price gap is enough to make them willing to take on the risk of a modified card.
Risks and Costs
The costs of modified graphics cards are also clear. First, there is the warranty issue: modified cards have no Nvidia manufacturer warranty at all, nor can continued driver updates be guaranteed. Second, there is uncertainty around stability and lifespan; nonstandard memory configurations and modified firmware may affect long-term stable operation. In addition, such modified cards are usually shipped through international logistics, the return and exchange process is cumbersome, and buyers must bear the risk of damage during transit.
The community’s attitude toward this kind of product is also quite conflicted. A Japanese developer on X said, “Taking on something like this is too risky, but everyone wants it,” while someone else shared their experience using a modded RTX 4090 48GB, saying, “It works perfectly and runs at high speed.”[7]The 96GB VRAM capacity is directly relevant to running large language models locally. For example, when running models at the 70B parameter level, 96GB of memory can greatly reduce reliance on multi-GPU parallelism or cloud inference, which is also the main appeal of modded cards.
Market Background: The Gray Industry Spawned by AI Demand
The rise of China’s modified graphics card supply chain stems from the gap between AI research’s hard requirement for large-capacity VRAM and the high pricing of Nvidia’s professional cards. Nvidia’s professional-grade 96GB graphics cards target enterprise customers, but their high prices push researchers and small companies to seek alternatives. In the European market, RTX 5090 prices have surged more than 70% since early this year, and the standard version’s price in the German market has exceeded 5,000 euros (about NT$178,000), making the cost-performance advantage of modified cards even more prominent.
In recent years, Chinese factories have produced several modified graphics cards, including the RTX 5080 32GB, RTX 4090 48GB, and RTX 3090 48GB, and these modifications mainly serve the AI market. The RTX 5090 96GB is the latest addition to this product line and the model with the largest memory expansion.
For developers considering this type of product, it is advisable to carefully assess practical issues such as firmware stability, driver compatibility, and after-sales support before placing an order. The $3,888 price is certainly tempting, but without a warranty and stable driver support, this card is suitable for advanced users with hardware experience who can troubleshoot issues themselves, not for general consumers. The continued expansion of the modded-card market also reflects how Nvidia’s pricing strategy between consumer and professional tiers is creating substantial room for third-party modders to operate.
Source: KOCPC Chinese