A workshop in Shenzhen listed an RTX 5090 modified to carry 96GB of video memory on Alibaba for roughly $3,888. The retail card ships with 32GB, so this is triple the memory — at around 35% less than a standard 5090 costs in the US.

Three times the memory for less money. That combination should sound wrong, and following why it is possible tells you something real about the current AI hardware market.

How you triple the memory

The answer is a new circuit board. According to reporting, the mod uses a custom PCB with a clamshell arrangement — memory populated on both sides of the substrate. Using both faces instead of one doubles what fits in the same area, and larger-capacity memory chips take it to triple.

The GPU die itself appears to be harvested from existing RTX 5090 cards and reballed onto the new board. To make the expanded memory usable, the firmware has to know about the configuration, which reportedly means leaked or modified firmware.

So this is not a new product. It is existing parts recombined by hand — which is exactly why there is no warranty and no guarantee of long-term stability.

Why memory capacity matters so much

To run an AI model at speed, the model has to fit in GPU memory. When it does not fit, data shuttles to system memory or storage, and throughput drops by multiples — sometimes by an order of magnitude.

So at identical compute performance, 32GB versus 96GB changes which models you can run at all. Slightly slower compute with the model resident beats faster compute with the model spilling.

That is where demand for these cards comes from. Nobody buys one for gaming. People buy them to fit a large model onto a single card.

Why these come out of China

High-end AI accelerators are supply-constrained and subject to export controls. When the capacity you want is hard to obtain through official channels, workaround demand appears.

There are two ways to meet it: use large quantities of whatever is permitted, or modify existing products to expand capacity — which is what is happening here. The fact that workshop-scale operations keep producing these is evidence the demand is real.

Why not to buy one

Cheap and high-capacity is appealing, but the risks here compound and none of them are verifiable.

• No warranty. Manufacturer coverage ends at the moment of modification. A failure is terminal.
• Unproven stability. A clamshell configuration differs from the original design in thermals and signal integrity. Behavior under sustained load is unknown.
• Driver and firmware risk. An official driver update can break a modified configuration. Not updating means not getting security patches either.
• No resale market. If something goes wrong, there is effectively nowhere to sell it.

Unless you have research budget you can afford to lose and can absorb outright failure, this is something to read about rather than buy.

What the news actually signals

More important than one card's specs is that it exists at all. Hand-built hardware selling at this price means demand for memory capacity is running ahead of supply.

Which makes the memory configuration of the next generation of consumer GPUs the thing to watch. If official products ship enough capacity, this cottage industry disappears. If they do not, it keeps going.