← All guides

How to Buy a Used RTX 4090 Safely (2026): It Costs $3,000

A used RTX 4090 costs about $3,000 in October 2026. That is 1.9 times its $1,599 launch price and almost twice a used RTX 3090. At that price, the 4090 also attracts scams the 3090 does not: boards with the GPU die and VRAM removed, China-market 4090D cards sold as 4090s, and 48GB modded cards with hacked firmware. Here is the price, then the checklist in the order that catches problems earliest.

Bottom Line

  • What it costs now: about $3,000-3,400 used, as of October 2026. That is 1.9 times the $1,599 launch price.
  • What you are buying: 24GB of GDDR6X at 1,008 GB/s. That is the same VRAM as a used RTX 3090, at almost twice the price.
  • The scam to know: repair shops have received 4090s with no GPU die and no VRAM on the board. The parts went to the China AI-mod market.
  • The 30-second check: GPU-Z must show “GeForce RTX 4090” and 16,384 shading units. 14,592 means a 4090D.
  • The test that matters: memtest_vulkan. Any error above zero, return the card.
  • Your walk-away price: about $3,300. Below $2,500, assume a problem until the tests prove otherwise.

First, the price

The 4090 does not lose value. NVIDIA reportedly stopped production around October 2024, before the RTX 5090 launch. NVIDIA did not announce this; supply-chain sources reported it. In 2026, the memory shortage pushed GPU prices up again, and used 4090s followed.

Prices checked on October 22-23, 2026:

CardSourceOctober 2026 figure
RTX 4090, usedResalePrices (9 listings, thin)$3,000 average and median; 90-day range $2,497-3,075
RTX 4090, marketplaceGPU Poet (90 listings, condition not split)From $3,399; October lowest average $3,147; daily range $2,997-3,300
RTX 3090, usedResalePrices (317 listings)$1,541 average; fair range $1,498-1,600
RTX 3090, marketplaceGPU Poet (83 listings)From $1,357; October lowest average $1,377
RTX 5090, marketplaceGPU Poet (92 listings)From $6,987

New 4090 stock is thin. Our last read was $3,200 on Amazon on September 25, 2026. We did not re-check it tonight. A new RTX 4090 with a warranty can cost only a few hundred dollars more than a used one. Compare both on the day you buy.

What a used 4090 costs per GB, and per GB/s

This is the number most buying guides skip. All figures below are derived from the October prices above and NVIDIA’s published specs.

CardVRAMBandwidthPrice used$ per GB VRAM$ per GB/s
RTX 309024GB936 GB/s$1,541$64$1.65
RTX 409024GB1,008 GB/s$3,000$125$2.98
RTX 509032GB1,792 GB/s$6,987 (lowest listing)$218$3.90

A used 4090 costs 1.95 times a used 3090 for the same 24GB and 8% more memory bandwidth. Bandwidth sets the ceiling on token generation speed. The 4090’s real advantage is compute: 16,384 CUDA cores against the 3090’s 10,496. That helps prompt processing, long contexts, and image generation. If you mostly chat with one model, you pay double for a small speed gain.

The checklist, in the order that catches problems earliest

1. Pay only where you can get your money back

Do this before you look at a single photo. eBay’s Money Back Guarantee covers an item that arrives broken, faulty, or not as described. You have 30 days after delivery to open a request, or the seller’s return window if it is longer. It does not cover items paid outside eBay checkout.

A seller who asks for Zelle, wire, or crypto removes your only protection. At $3,000, walk away.

2. Check the listing for “no core” and “for parts” cards

Search eBay for “RTX 4090 no core” and you will find real listings: ASUS TUF, ROG Strix, and PNY 4090s sold “for parts” with no GPU core and no VRAM. They include the box, cooler, fans, and backplate. They look like a complete card in photos.

The source of these boards is the China AI market. Modders remove the 4090’s GPU and memory and solder them onto new PCBs, often with more VRAM. The empty donor boards then come back to Western marketplaces.

Some of them get sold as working cards:

  • Tom’s Hardware reported an ASUS TUF RTX 4090 bought used on eBay that would not post. The repair shop Northridge Fix opened it and found no GPU core and no VRAM chips on the PCB.
  • Wccftech reported (April 29, 2024) a 4090 bought on Facebook Marketplace with the GPU and memory removed. Only the PCB was left.
  • TechSpot covered a repair technician who found three of four “RTX 4090s” were fakes with RTX 3000-series GPUs inside.

Rule: if a card does not post, return it at once. Do not let a seller talk you into “it needs a driver reinstall.”

3. Confirm it is a 4090, not a 4090D

The RTX 4090D is the China-market version that NVIDIA released in December 2023 to comply with US export rules. It has the same 24GB of GDDR6X and the same 1,008 GB/s bandwidth. But it has fewer cores:

SpecRTX 4090RTX 4090D
CUDA cores16,38414,592
Tensor cores512456
RT cores128114
Power (TGP)450W425W
OverclockingStandardRestricted

The 4090D has a different device ID. GPU-Z reports it as “RTX 4090 D” and shows 14,592 shading units. A real 4090 shows 16,384. Ask the seller for a GPU-Z screenshot before you pay, and check it again when the card arrives. On eBay, 4090D listings, including 48GB versions, mostly ship from Shenzhen.

The 4090D is not a bad card. But it has about 11% fewer CUDA cores (derived: 16,384 / 14,592). Do not pay 4090 money for it.

The same check catches the fakes in step 2. A card with a swapped older GPU shows the wrong name and core count in GPU-Z.

4. Know what a 48GB 4090 is before you buy one

You will see “RTX 4090 48GB” listings. NVIDIA never made this card. Modders move the 4090 GPU to a new PCB that holds 24 GDDR6X chips of 2GB each, on both sides, then flash a leaked, modified firmware.

48GB on one card is attractive for 70B models. But no board partner built it, so there is no manufacturer warranty to claim. Buy one only from a seller who accepts returns, and run every test below on it.

5. Inspect the 16-pin power socket

The 4090 uses the 16-pin 12VHPWR connector. In November 2022, NVIDIA said it had 50 known reports of melted connectors, and named improper insertion as the likely main cause. One engineer showed that a gap of only 1mm increases resistance.

Look at the socket on the card:

  • Discolored, browned, or warped plastic.
  • Bent or blackened pins.
  • Damage that starts from one side of the connector. Gamers Nexus found this pattern points to a plug that was not fully inserted.

A working card does not prove the socket is clean. Tom’s Hardware reported a reviewer’s official 4090 power cable that had melted, hidden, while the card kept working for two years. Ask for a clear photo of the socket before you buy.

6. Run memtest_vulkan

Run memtest_vulkan. Any error count above zero means return the card.

It uses Vulkan compute, so it needs no CUDA install. It runs on Windows and Linux. The author says to wait at least 6 minutes and that more than 2 hours is usually unneeded. We run it for 30 minutes on a card that costs $3,000.

7. Load it, and watch memory temperature

Run a long, heavy load: a big llama.cpp model that fills most of the 24GB, for 30 minutes or more. Watch these values:

  • In nvidia-smi: the name, total memory (about 24GB), power draw, and core temperature. nvidia-smi --query-gpu=name,memory.total,power.draw,temperature.gpu --format=csv -l 5 prints them every 5 seconds.
  • In HWiNFO or GPU-Z: Memory Junction Temperature. GDDR6X runs much hotter than the core. Igor’s Lab measured GDDR6X above 100°C on an RTX 3080 FE. Compare the memory reading with the core reading. A memory temperature near or above 100°C under load is a reason to ask about the thermal pads.

Watch for crashes, driver resets, or garbage output. If the card survives a long run with real inference, it is a good card.

When the used 4090 is the wrong buy

Skip it if you only need 24GB. A used RTX 3090 gives you the same 24GB for about half the price. How to buy a used RTX 3090 safely has that checklist.

Skip it if you want a 70B model. 24GB does not hold a 70B at a useful quant. Can an RTX 4090 run a 70B? works through why.

Skip it if the price is close to new. A new card with a warranty removes every risk on this page. When the used price is within a few hundred dollars of new stock, buy new.

Skip it if you plan to resell soon and bought at the peak. The 4090 held its value so far. Local AI rig resale value covers what usually happens when the next generation gets cheaper.

What it runs once you have it

24GB is the tier where 27-35B models fit, and the 4090’s compute makes long prompts fast. Best local LLM for the RTX 4090 has the current model picks and quant sizes. Read it once the card passes memtest.

Sources

See Also

Read next

Is There a 48GB RTX 5090? No: What to Buy Instead
NVIDIA sells no 48GB RTX 5090 and no 48GB GeForce card. The '5090 48GB' in the news is a modding prediction, not a product. Here are the real 48GB routes as of September 2026, and why a used 48GB workstation card now costs less than a stock 32GB 5090.
Two Used RTX 3090s or One RTX 5090? 48GB Slow vs 32GB Fast
Dual used RTX 3090s cost $2,000-2,600 for 48GB of VRAM. One RTX 5090 costs $4,300-5,000 for 32GB. The 2026 price spike flipped this comparison: the dual build is now half the price AND holds a 70B. Here is the honest tradeoff, including the 700W problem.
Best Local LLM for RTX PRO 5000 (2026): gpt-oss 120B Wins
What runs on the RTX PRO 5000 Blackwell 48GB and 72GB: gpt-oss 120B with full 128K context on the 72GB, Qwen3.6 35B-A3B at Q8 on the 48GB, measured tok/s, $/GB.
Best LLM for 64GB VRAM GPUs (2026): Laguna S 2.1 Wins
No consumer GPU has 64GB of VRAM. The cards that get you there, and the best LLM for them: Laguna S 2.1 UD-IQ4_XS (53.61 GiB) over two 32GB cards.