Modded RTX 5090 with 96GB of VRAM Listed in China

Yes, the listing is real. No, $900,000 is not the price. And no, nobody has proven the card actually works. A modified GeForce RTX 5090 with 96GB of video memory — three times the stock card's 32GB — has been listed on Xianyu, Alibaba's secondhand marketplace, by a seller in Changsha, China. The listing was spotted by hardware repair specialist VIK-on, first reported by VideoCardz, and covered by WCCFTech on 12 November 2026. Everything after that first sentence needs careful handling, because the listing is a business-to-business offer with a placeholder price, a confidentiality agreement, and no independent benchmark anywhere in public.

This is not an Nvidia product. Nvidia has never sold a 96GB GeForce RTX 5090, and the company has not commented on the listing. What the seller is offering is a transplant: the GB202 graphics processor from a retail RTX 5090, removed from its original board and soldered onto a custom printed circuit board with far more memory than Nvidia ever shipped on a consumer card. If it works as claimed, it would match the memory of Nvidia's RTX PRO 6000 Blackwell workstation card — a card that lists at around $16,000 — at a privately negotiated price almost certainly far below the headline number. That "if" is doing a lot of work, and this article keeps it visible throughout.

What the listing actually says

The listing appeared on Xianyu, which is Alibaba's secondhand marketplace and is also known as Goofish. It was noticed by VIK-on, a hardware repair specialist who tracks the Chinese GPU modification scene, and first reported in English by VideoCardz before WCCFTech and Notebookcheck picked it up on 12 November 2026. The seller is based in Changsha, the capital of Hunan province, and operates under the handle "Super cute Minami Kotori".

What is on offer is not just a graphics card. The package, as described in the coverage, is a complete modification kit: the modified 96GB card itself, the custom PCB schematics, and the patched BIOS firmware needed to make the card recognise and address the extra memory. That last part matters more than it sounds. A graphics card's BIOS tells the GPU how much memory exists and how to talk to it; without modified firmware, extra memory chips soldered onto a board are invisible to the system. The seller is, in effect, selling the know-how as well as the hardware — which is one reason the offer is aimed at businesses rather than individual buyers.

The commercial terms are unusually strict for a marketplace listing, and they are part of the story. Individual buyers are explicitly discouraged from making inquiries. Corporate buyers must identify their company and sign a confidentiality agreement before any transaction moves forward, and hardware testing is available only through verified intermediaries. WCCFTech's reporting frames these barriers as protection against regulatory scrutiny and intellectual property leaks — a plausible reading, since the seller is advertising a product built from Nvidia's flagship consumer silicon, modified in ways Nvidia never authorised, for a market Nvidia's export-controlled product line does not officially serve.

The headline price is 6,488,888 yuan. That number is a placeholder, and the listing says so: the actual price is negotiated privately with qualified buyers. Placeholder prices with long strings of lucky digits are common on Chinese marketplaces — 8 is considered a lucky number, and 6,488,888 is a deliberately eye-catching figure rather than a serious asking price. WCCFTech converts it to approximately $900,000 and notes it is about 57 times the RTX PRO 6000's $16,000 figure. The arithmetic is worth stating plainly: $900,000 divided by $16,000 is 56.25, so "57 times" is a rounded headline number. The conversion itself is approximate too — at the exchange rate on 15 November 2026, about 6.7 yuan to the dollar, 6,488,888 yuan is roughly $967,000, not $900,000. Neither figure is the price of the card. The only honest statement about the price is that it is unknown and negotiated behind closed doors.

How the mod works — and why a stock 5090 cannot just take more memory

To understand the modification, start with what Nvidia actually ships. The GeForce RTX 5090 launched on 30 January 2025 at $1,999, built on the GB202 graphics processor with 32GB of GDDR7 memory on a 512-bit memory bus. The 512-bit bus is the key number: it means the GPU talks to memory through 16 separate 32-bit channels, and each channel is wired to one memory chip on the board. The stock card uses sixteen 2GB GDDR7 chips, one per channel, for a total of 32GB. There are no empty memory slots on a graphics card, and no spare channels. You cannot plug in more memory the way you add RAM sticks to a motherboard.

The mod gets around this with a technique called clamshell mounting. Instead of one chip per channel, the board carries two chips per channel — one on the front of the board and one on the back, sharing the same memory channel. The board in the listing photos shows exactly this: Samsung memory modules packed densely on both sides of a custom PCB around the central GB202 die. The modules are 3GB GDDR7 chips, according to VideoCardz's reporting of the listing. Thirty-two 3GB chips — sixteen per side — add up to 96GB. The stock card's sixteen 2GB chips add up to 32GB. Same GPU, same 512-bit bus, three times the memory.

That is the easy half of the job. The hard half is everything else. A custom PCB has to be designed and manufactured to route twice as many memory chips, with signal integrity good enough for GDDR7 speeds, and with power delivery capable of feeding the extra chips. The GB202 die itself has to be harvested from a donor RTX 5090 — desoldered from its original board and reballed onto the new one — which is precision rework very few workshops in the world can do reliably. Then the firmware has to be patched so the card's BIOS, and Nvidia's drivers, recognise 96GB instead of 32GB. That firmware step is where earlier attempts reportedly stalled: VideoCardz noted in October 2026 that there was still no publicly verified method for upgrading a 5090's memory, despite claims from Chinese suppliers about 64GB and 96GB versions. A US repair technician, Greg Sky of GPU Lab in Michigan, posted on 31 October 2026 that "64 and 96gb 5090 confirmed" and that 5090 firmware had been unlocked — but he shared no photos, no benchmarks, and no firmware details, and only three days earlier he had posted that there was no 96GB 5090. A claim that flips in under a week carries less weight than a repeatable measurement.

There is also a bandwidth question the 96GB headline hides. Adding memory chips in clamshell does not widen the memory bus — it is still 512 bits wide — so the card's memory bandwidth stays at the stock card's level of about 1,792 GB/s. What triples is capacity, not speed. For the AI workloads this card is aimed at, that is the right trade: the bottleneck is usually fitting a model's weights into memory at all, not how fast the memory is.

Stock 5090 vs modded 5090 vs RTX PRO 6000

The comparison that makes the listing interesting is not against other gaming cards. It is against Nvidia's own professional card built on the same silicon. The RTX PRO 6000 Blackwell Workstation Edition uses the same GB202 die, carries 96GB of GDDR7 memory with error correction, and lists at $16,000 on Nvidia's own US marketplace — a price Nvidia doubled from the $8,565 launch figure in August 2026. The modded card, if it works, offers the same memory capacity on the same die, without error correction, without Nvidia's warranty, and without Nvidia's professional driver support, at a privately negotiated price. Whether that trade is worth it depends entirely on the price nobody has published.

Specification	GeForce RTX 5090 (stock)	Modded RTX 5090 (listed)	RTX PRO 6000 Blackwell
GPU	GB202	GB202 (harvested)	GB202
VRAM	32GB GDDR7	96GB GDDR7 (clamshell)	96GB GDDR7 with ECC
Memory bus	512-bit	512-bit	512-bit
Memory bandwidth	~1,792 GB/s	~1,792 GB/s (unchanged)	1,792 GB/s
CUDA cores	21,760	21,760 (same die)	24,064
Power draw	575W	Higher (extra memory chips)	600W
Launch / list price	$1,999 (30 Jan 2025)	Placeholder 6,488,888 yuan (~$900,000 per WCCFTech)	$16,000 (Nvidia US marketplace)
Warranty	Nvidia / board partner, 3 years	None	Nvidia professional warranty
Status	Shipping product	Unverified listing, no benchmarks	Shipping product

Two rows in that table deserve a second look. The modded card has fewer CUDA cores than the PRO 6000 — 21,760 against 24,064 — because the consumer GB202 die ships with some cores disabled, and the mod changes the memory, not the silicon. And the modded card's power draw is listed as "higher" with no number, because no number has been published: thirty-two memory chips draw more power than sixteen, and the custom board's power delivery is unrated in public.

Why anyone would want 96GB of VRAM

For gaming, 96GB of VRAM is almost meaningless. No game released in 2026 needs anything close to it, and a stock RTX 5090's 32GB already covers the most demanding gaming workloads with room to spare. The buyers for a card like this are not gamers. They are people running artificial intelligence workloads locally — and for them, VRAM is the wall that decides what is possible at all.

Running a large language model locally means loading the model's weights into the graphics card's memory. If the weights do not fit, the model does not run — or it runs partly in system RAM, which is so slow that generation crawls. A 70-billion-parameter model in 8-bit precision needs roughly 70GB for its weights alone, before any space for conversation context. That is more than double a stock 5090's 32GB, and it fits comfortably in 96GB. A 120-billion-parameter model needs around 80GB in efficient precision — again, impossible on 32GB and comfortable on 96GB. The modded card's pitch is simple: one card, one PCIe slot, models that otherwise need two or three cards, or a datacenter GPU you cannot buy.

That is also why the Chinese market specifically wants these cards. United States export controls restrict China's access to high-end datacenter GPUs, and Nvidia's China-specific consumer card, the RTX 5090D, ships with reduced AI performance to comply. Workshops in China have responded by rebuilding the desktop cards they can obtain: harvesting the dies, designing custom boards, and tripling or quadrupling the memory for AI server use. VideoCardz reported in October 2026 on a large Chinese workshop — described as more of a production line than a modification shop — where retail RTX 5090 cards, including Founders Edition models, were being stripped down and rebuilt for AI servers. The 96GB listing is the retail-facing end of that same pipeline.

There is a useful sense of scale here from our own coverage. A stock RTX 5090's 32GB already covers demanding local work — for example, native 4K video generation with LTX-2 needs around 24GB of VRAM, which fits inside the stock card with room to spare. The 96GB card is not for that kind of workload. It is for the workloads that do not fit in 32GB at all: the largest open models, long-context inference, and fine-tuning runs where the model, the optimizer state, and the activations all have to live in memory together. For those users, the question is never "is 96GB faster" — it is "does it fit".

The September listing: a different card, a different story

This is not the first 96GB RTX 5090 to appear on a Chinese marketplace, and confusing the two listings is the easiest mistake in this story. In September 2026, Shenzhen Suqiao Intelligent Technology listed a modified RTX 5090 with 96GB of VRAM on Alibaba for $3,888 — a price that Tom's Hardware noted was about 35% below the US price of a standard RTX 5090 at the time. That listing had a problem the November listing does not: its own specifications contradicted themselves.

The September listing described the memory as GDDR6X running at 14Gbps. The retail RTX 5090 uses GDDR7, not GDDR6X, and 14Gbps is far below GDDR7's operating speeds. As Guru3D pointed out, GDDR6X was only ever made in 2GB chips, so even in a clamshell layout the maximum capacity would be 64GB — 32 chips of 2GB each — not 96GB. The numbers in the listing could not produce the capacity the listing claimed. Tom's Hardware called the specifications "not entirely accurate", and Igor's Lab reached the same conclusion: until an independent test of a delivered card exists, it is impossible to know which components are actually installed.

The September listing then vanished. It could no longer be found on Alibaba, and at least one outlet that covered it added an editor's note saying the information could not be found again and that it may have been a gimmick. That is a very different profile from the November listing, which names a specific memory type that matches the hardware — Samsung 3GB GDDR7 chips, which do add up to 96GB in clamshell — and which comes with a seller offering testing through intermediaries rather than a take-it-or-leave-it storefront price. The September listing is evidence that this market produces listings that do not add up. The November listing is evidence that the mod itself is getting more plausible. Neither one is evidence that a working 96GB card exists.

The verification problem

Everything in this story so far is a claim about a listing. What is missing is the one thing that would settle it: proof that a modified card works. As of 15 November 2026, no independent benchmark of a 96GB RTX 5090 has been published. No teardown has been published. No GPU-Z screenshot — the standard utility that reads and reports a card's specifications — has been published. No video shows the card running a model, a game, or even a Windows desktop. The listing photos show a bare board, and a bare board is not a working card.

That absence matters more than it might seem, because the failure modes of this kind of modification are invisible in a photo. A card can report 96GB in software while only 32GB is actually addressable. A card can boot and idle while crashing under load because the memory signal integrity falls apart at GDDR7 speeds, or fail after an hour as densely packed memory chips overheat. None of these can be checked from a marketplace photo, and the seller's terms — testing only through verified intermediaries, under a confidentiality agreement — mean that even a successful private test may never become public evidence.

What would prove it? In order of strength: a GPU-Z screenshot from a named, independent tester showing 96GB of GDDR7 on a 512-bit bus; a memory bandwidth test showing the full ~1,792 GB/s across the whole capacity; a sustained load test, such as loading a 70B model and generating text for an hour without errors or crashes; and a teardown confirming Samsung 3GB GDDR7 chips on both sides of the board. The strongest proof would be all four from the same card, in the same hands, on camera. Until something like that appears, the correct description of this card is "claimed" or "alleged" — the word VideoCardz itself used in its headline.

The risks, stated plainly

No warranty. Modifying a graphics card voids its warranty — Nvidia's GeForce warranty runs three years from purchase and excludes damage and misapplication by a non-authorised party, and harvesting the die onto a different board is about as far from authorised service as it is possible to get. If the modded card fails, there is nobody to send it back to.

Not an Nvidia product. Nvidia did not make this card, did not approve it, and does not support it. It will not appear in Nvidia's driver support as a distinct product, and any driver or firmware update could, in principle, break the patched BIOS the card depends on. The buyer would be depending on an unnamed seller's firmware patch for the working life of a very expensive piece of hardware.

The confidentiality agreement cuts both ways. A buyer who signs a non-disclosure agreement before testing cannot publish the results of that test. That is convenient for a seller whose product might not pass, and it is a structural reason why independent verification may be slow to appear even if testing does happen. A test nobody is allowed to talk about is not verification.

The scam risk is real. The September listing shows that this market produces offers whose specifications do not add up, and the November listing's own terms — huge placeholder price, private negotiation, intermediaries only — are exactly the shape of a transaction that is hard to reverse if it goes wrong. None of that proves the November seller is dishonest. It proves that a buyer has no cheap way to find out, and that the burden of proof is entirely on the seller.

The export-control context. These modifications exist largely because of US export controls on high-end AI silicon to China, and Nvidia's China-specific RTX 5090D is the compliant alternative the controls permit. A business buying a modified card is buying into a grey area between a consumer product and a controlled technology, and the rules around that area change. That is a legal and compliance question for the buyer's own advisers, not something a marketplace listing answers.

What happens next

Three things are worth watching. First, whether any verified intermediary publishes a test — a GPU-Z screenshot, a bandwidth result, or a sustained model run would change this story from a listing into a product overnight. Second, whether the seller's real price leaks or is reported: the gap between the placeholder and the real price will say a lot about who the intended buyer is. A price near the PRO 6000's $16,000 would mean the card is competing with Nvidia's professional card on capacity alone; a price far below it would mean the seller is undercutting on the strength of harvested silicon and no warranty. Third, whether other workshops follow with their own 96GB boards. The September listing, the October factory report, and the November listing together describe a capability spreading through an industry, not a single seller's one-off. If the firmware unlock Greg Sky described is real and repeatable, 64GB and 96GB 5090s could become a small but real category — and verification would become the whole story.

Until then, the honest summary fits in one paragraph. A real listing advertises a modified RTX 5090 with 96GB of VRAM on a Chinese secondhand marketplace. The price in the listing is a placeholder, the seller will only deal with businesses under confidentiality, and nobody outside the seller's circle has shown the card working. The modification is technically plausible — the memory math adds up, the chips exist, and the workshops exist — but plausible is not proven, and in a market where the last listing's specifications contradicted themselves, proof is the only currency that counts.

Verdict

The listing is real, the price is not, and the card is unproven. That is the whole story, and it is a more interesting story than the headline number. A $900,000 graphics card would be a curiosity. A working 96GB RTX 5090 at a plausible price would be a genuine event for local AI — the card that lets one PCIe slot hold models that currently need a datacenter GPU. The gap between those two stories is entirely made of verification, and it has not been closed yet. Watch for the benchmark, not the price tag.

Frequently Asked Questions

Is the $900,000 price real?

No. The listing price is 6,488,888 yuan, which WCCFTech converts to approximately $900,000, and it is explicitly a placeholder. The seller says the real price is negotiated privately with qualified business buyers. At the 15 November 2026 exchange rate the same yuan figure is roughly $967,000 — the conversion is approximate, and neither dollar figure is what anyone would pay.

Can I buy one?

Almost certainly not as an individual. The seller targets business-to-business clients only, requires corporate buyers to identify their company and sign a confidentiality agreement, and explicitly discourages inquiries from individual buyers. Hardware testing is available only through verified intermediaries. There is no published price, no warranty, and no return policy.

Does the card actually work?

Nobody outside the seller's circle has proven it. No independent benchmark, teardown, or GPU-Z screenshot of a working 96GB RTX 5090 has been published as of 15 November 2026. The modification is technically plausible — Samsung 3GB GDDR7 chips in a clamshell layout do add up to 96GB — but plausible is not proven, and the card should be described as claimed or alleged until a verified test appears.

Is this an Nvidia product?

No. Nvidia has never sold a 96GB GeForce RTX 5090 and has not commented on the listing. The card is a third-party modification: a GB202 die harvested from a retail RTX 5090 and mounted on a custom PCB with modified BIOS firmware. It carries no Nvidia warranty and no Nvidia support.

Why would anyone want 96GB of VRAM?

For local AI work. Large language models have to fit their weights into VRAM, and a 70-billion-parameter model needs roughly 70GB while a 120-billion-parameter model needs around 80GB — both impossible on a stock 5090's 32GB and comfortable on 96GB. For gaming, 96GB is almost meaningless; no current game needs anything close to it.

What about the September listing for $3,888?

That was a different, earlier listing from a different seller — Shenzhen Suqiao Intelligent Technology on Alibaba — and it should not be confused with the November Xianyu listing. The September listing described the memory as GDDR6X at 14Gbps, which contradicts the RTX 5090's GDDR7 hardware and could not add up to 96GB. The listing later disappeared from Alibaba. It is a warning about this market, not evidence about the November card.

Sources

• WCCFTech — "Modded RTX 5090 With 96GB VRAM Shows Up On Chinese Marketplace With A Placeholder Price That's 57x Higher Than The RTX PRO 6000's" (12 November 2026): https://wccftech.com/modded-rtx-5090-96gb-vram-china-listing/
• Notebookcheck — "Modded Nvidia RTX 5090 with 96 GB VRAM and custom BIOS surfaces in China" (12 November 2026): https://www.notebookcheck.net/Modded-Nvidia-RTX-5090-with-96-GB-VRAM-and-custom-BIOS-surfaces-in-China.1448783.0.html
• Tom's Hardware — "China-modified Nvidia RTX 5090 with massive 96GB of memory appears on Alibaba for less than $4,000" (11 September 2026, Zhiye Liu): https://www.tomshardware.com/pc-components/gpus/china-modified-nvidia-rtx-5090-with-massive-96gb-of-memory-appears-on-alibaba-for-less-than-usd4-000-3x-more-vram-at-65-percent-the-cost-of-the-original
• Igor's Lab — "RTX 5090 with 96 GB VRAM: Chinese Modification Triples Memory" (September 2026): https://www.igorslab.de/en/rtx-5090-96-gb-vram-chinese-modification-triples-nvidia-graphics-memory/
• ShiftDelete — "RTX 5090 96GB Mods Confirmed by US Modder: What We Know" (November 2026, Greg Sky / GPU Lab firmware-unlock claim): https://shiftdelete.net/en/rtx-5090-96gb
• Nvidia — GeForce RTX 5090 product page (32GB GDDR7): https://www.nvidia.com/en-us/geforce/graphics-cards/50-series/rtx-5090/
• Nvidia — RTX PRO 6000 Blackwell Workstation Edition product page (96GB GDDR7 with ECC, 1,792 GB/s, 600W): https://www.nvidia.com/en-us/products/workstations/professional-desktop-gpus/rtx-pro-6000/
• Run AI Home — "RTX PRO 6000 Blackwell for Local AI in 2026" (Nvidia US marketplace $16,000 listing; price doubled from $8,565 launch MSRP in August 2026): https://www.runaihome.com/blog/rtx-pro-6000-blackwell-local-ai-2026/
• VideoCardz — original report of the Xianyu listing (credited as the first reporter by WCCFTech and Notebookcheck; direct article URL could not be established from this VM — videocardz.com returns 403 to this machine and its search index did not surface the article URL, so no direct link is given): cited via WCCFTech and Notebookcheck above.

Related reading: Jensen Huang resigns from Nvidia · gputaskman: the free GPU task manager · Colibri: run 744B AI models on a potato PC