RTX 4090 48GB: how the modded cards are made, what they cost, and whether to buy one

The RTX 4090 48GB is not an NVIDIA product. Workshops in China move a 4090's AD102 chip onto a custom double-sided board with 24 GDDR6X chips of 2 GB and flash modified firmware, as Tom's Hardware described in September 2025. On 7 November 2026 the reseller MillionMiner listed one at $4,099, in stock, and Alibaba supplier listings ran from about $3,999 to $5,066. It keeps the stock card's 1,008 GB/s, so speed is a normal 4090's; the gain is capacity, enough for Llama 3.3 70B at 4-bit with about 23,000 tokens of context by my calculation. It comes with no NVIDIA warranty, a loud blower cooler, and a fake problem: some cards only report 48 GB.

Dollar to euro at the ECB reference rate of 1.1165 for 6 November 2026; my conversion.

How is a 48 GB RTX 4090 made?

By moving the GPU to a different board. The GeForce RTX 4090 has memory on one side of its PCB only, while NVIDIA's professional cards put chips on both sides ("clamshell") to double capacity on the same bus. According to Tom's Hardware, citing Russian technician VIK-on, Chinese suppliers sell a kit for about $142: a pre-populated clamshell PCB plus a blower cooler.

1. Desolder the AD102 chip and its twelve 2 GB GDDR6X chips from a donor RTX 4090.
2. Solder them onto the clamshell board with twelve more GDDR6X chips, about $24 each on Chinese marketplaces.
3. Flash a leaked, modified firmware that accepts 48 GB.

Tom's puts the parts at about $430 beyond the donor card. The same firmware trick produced RTX 4080 Super 32 GB cards. The reason the business exists is US export policy: NVIDIA's 8-K filing of 17 October 2023 lists the RTX 4090 among chips that need a US license for export to China and to Country Groups D1, D4 and D5. A vLLM contributor describes the RTX 4090D, the China-market variant, as "the most widely adopted 48GB AI accelerator in the Chinese market" (PR #47732, July 2026).

Where are 48 GB RTX 4090s sold, and what do they cost?

Mostly by Chinese suppliers and a few Western resellers, at $3,750 to about $5,000 in 2026. These are listing prices, not sale prices.

For comparison on the same day, the cheapest new 24 GB RTX 4090 on Geizhals was €4,989.99 from a marketplace seller, and the cheapest RTX 5090 listing was €6,249 (Geizhals), both incl. 19% VAT. The Alibaba page also mixes in 24 GB cards, RTX 6000 Ada and RTX 5090 listings under the same search, so read each title. Current RTX 5090 prices are tracked in the RTX 5090 price article.

The odd result: a modded 48 GB 4090 at about €3,671 now costs less than a new 24 GB 4090 in Germany. New 24 GB cards are scarce: Geizhals showed a single offer for the Founders Edition.

Is a 48 GB RTX 4090 faster than a normal 4090?

No. Bandwidth decides decode speed, and the mod does not change the 384-bit bus or the 21 Gbps memory. igor'sLAB measured gaming and synthetic results no better than a stock card. For LLMs, the llama.cpp CUDA scoreboard gives a useful bracket on Llama 2 7B Q4_0 with flash attention: 189.0 tok/s decode for a stock RTX 4090 and 179.5 tok/s for the RTX 6000 Ada, NVIDIA's own 48 GB AD102 card with slower GDDR6. I found no scoreboard entry for a modded card.

The software caveats are the Ada ones, not the mod's. Ada (sm_89) has FP8 but no FP4 tensor cores (Geizhals lists 660.5 dense FP8 TOPS and no FP4 figure), so NVFP4 checkpoints do not get native kernels. An owner of two 48 GB cards hit "Sinks are only supported in FlashAttention 3" running gpt-oss in a vLLM image in August 2025 (vLLM #22331). The cards do run current stacks: an SGLang contributor listed "RTX 4090 48GB mod (SM89)" as test hardware in July 2026 (SGLang #31190). For format support by GPU generation, see the quantization guide.

What does 48 GB let you run that 24 GB cannot?

A 70B model at 4-bit, or 27B to 35B models at 8-bit with long context. I compared weights plus KV cache plus 1.5 GB of overhead (my assumption) against 24 GiB and 48 GiB.

Llama 3.3 70B: 80 layers, 8 KV heads, head dim 128
KV per token = 80 x 2 x 8 x 128 x 2 bytes = 327,680 bytes at FP16
48 GiB = 51.54 GB; free = 51.54 - 1.5 - 42.52 (Q4_K_M) = 7.52 GB
context = 7.52 GB / 327,680 B = about 23,000 tokens (about 43,000 with q8_0 KV)
decode ceiling = 1,008 GB/s / 42.52 GB = about 24 tok/s

Weights from Hugging Face GGUF listings (Llama 3.3 70B, unsloth Qwen repos, ggml-org gpt-oss-120b), checked 7 November 2026. My calculations, not measurements. The 70B row is the reason most buyers want this card; how that model behaves on 128 GB is on the Llama 3.3 70B page.

What are the risks of buying a modded RTX 4090?

The card can be fine, but you carry every risk yourself.

• No NVIDIA warranty. MillionMiner's listing says the card is "not a factory NVIDIA product" and "No NVIDIA warranty applies"; any warranty is the seller's own terms. Its page also contradicts itself on this point in another section, which tells you how carefully to read it.
• Firmware. The cards run modified firmware. The seller says standard NVIDIA drivers see all 48 GB. Whether a future driver changes that is unknown; I found no report either way.
• Fakes. Alibaba's own listing page warns about "BIOS-only" cards that report 48 GB without the extra chips. Load a model larger than 24 GB, or run a full-memory test, on arrival.
• Noise and heat. igor'sLAB measured up to 65 dB from the blower and peaks of 86 °C. The listing quotes a 450 W TDP. Plan for a server room or a closed office door.
• Import and resale. Buying from outside the EU means customs clearance and import VAT unless the seller ships with duties paid (MillionMiner says it does). The card is a controlled item for China and the D1, D4 and D5 country groups under the US rule NVIDIA disclosed, so shipping one onward to those destinations is the case to check with a trade lawyer. This is not legal advice.

Is there an RTX 5090 with 96 GB or 128 GB?

Not one I could verify on 7 November 2026. The arithmetic shows why a modded 5090 is harder: its 512-bit bus has 16 channels, so 16 chips of 2 GB give 32 GB, 16 chips of 3 GB give 48 GB, and 32 chips of 3 GB in clamshell give 96 GB. That last layout is NVIDIA's RTX PRO 6000 Blackwell, which uses the same GB202 chip (Geizhals chip data). Tom's Hardware mentioned an RTX 5090 128GB rumor in September 2025 and said it "may seem like a scam now". BenchLife, via VideoCardz in October 2026, reports that NVIDIA is steering GB202 supply toward RTX PRO cards, which makes donor chips scarcer still. Treat any 96 GB or 128 GB "5090" listing as unproven until someone independent tests one.

What are the safer alternatives to a 48 GB RTX 4090?

Pay more for a warranty, or rent the capacity for the weeks you need it.

Geizhals prices include 19% VAT. The RTX 5090 bandwidth figure is from its Geizhals spec listing; I assume the RTX PRO 6000's from the shared GB202 chip and 512-bit bus. A 48 GB L40S rents for $0.79 an hour on the L40S page.

The RTX PRO 5000 costs about 2.3 times as much as the mod for the same 48 GB, and buys ECC memory, 33% more bandwidth, 300 W instead of 450 W and a real warranty. The RTX PRO 6000 costs about the same per gigabyte as the RTX PRO 5000 and doubles the capacity; the DGX Spark vs RTX PRO 6000 comparison covers that trade. For 70B to 120B models at low speed, a DGX Spark holds more than any of them.

My verdict: for a hobbyist who wants one 70B model on one card and can test it on arrival, the mod is the cheapest 48 GB on the market, and current inference stacks run on it. For a company, buy the RTX PRO 5000, or rent: GPUwerk's Spark and RTX 5090 bill per minute, and a full month of round-the-clock Spark rental ($576.70 at $0.79 per hour) is about a seventh of the modded card's price. For more on the stock card, see RTX 4090 vs RTX 5090 for LLMs, and for the 96 GB card's variants, RTX PRO 6000 Max-Q vs Workstation.

FAQ

Sources

Every figure above is attributed to its source in the sentence or table where it appears. Pages were read on 7 November 2026 unless a date says otherwise.

• $142 upgrade kit and spare modules turn Nvidia RTX 4090 24GB to 48GB AI card (Tom's Hardware, syndicated), Tom's Hardware via Yahoo Tech, 2025-09-10
• NVIDIA's modded GeForce RTX 4090 with 48 GB VRAM: test shows high background noise and disappointing gaming performance, igor'sLAB, 2025-03-28
• Chinese vendor sells water-cooled RTX 4090 with 48 GB VRAM, Wccftech, 2025-03-22
• NVIDIA RTX 4090 48GB Blower, MillionMiner
• 48GB graphics card 4090 listings, Alibaba.com (ElectronicsHub), 2026-08-15
• NVIDIA Form 8-K, 17 October 2023 (export licensing requirements), U.S. SEC EDGAR, 2023-10-17
• Performance of llama.cpp on Nvidia CUDA (scoreboard), ggml-org/llama.cpp on GitHub
• [Bug]: Sinks are only supported in FlashAttention 3 (4090 48gb), vllm-project/vllm on GitHub, 2025-08-06
• Add FP8 W8A8 Block kernel configs for NVIDIA GeForce RTX 4090D, vllm-project/vllm on GitHub, 2026-07-06
• feat(dsa): add SM89 FP8 paged indexer operator, sgl-project/sglang on GitHub, 2026-07-14
• NVIDIA RTX 5080 24GB reportedly launching in Q1 2027 (GB202 supply moving to RTX PRO), VideoCardz, 2026-10-12
• NVIDIA GeForce RTX 4090 Founders Edition price comparison, Geizhals
• NVIDIA RTX PRO 5000 Blackwell 48GB price comparison, Geizhals
• PRO 6000 Blackwell search, sorted by price, Geizhals
• INNO3D GeForce RTX 5090 X3 price comparison and chip data, Geizhals
• unsloth/Llama-3.3-70B-Instruct-GGUF file listing, Hugging Face
• ECB euro foreign exchange reference rates, European Central Bank, 2026-11-06