Used GPU Buying Guide 2026: Prices, Checks, What's Worth It

The new-card shelf sets every used price beneath it. The Tom's Hardware GPU price index, updated September 3, 2026, lists the best US price against launch MSRP: RTX 5090 $5,199 against $1,999 (+160%, our arithmetic); RTX 5070 $799 against $549 (+46%); RTX 5060 Ti 16 GB $758 against $429 (+77%); the out-of-production RTX 4070 $819 against $549 (+49%); Radeon RX 9060 XT 16 GB $469 against $349 (+34%); Intel Arc B580 $309 against $250 (+24%). Across Blackwell, per Tom's September 1 report, the average premium is “about 60% or 1.6x higher compared to MSRP.”

Part of it is on the record. CFO Colette Kress on the August 26 call: “we are experiencing extreme pricing conditions in memory. The magnitude of the price increase has exceeded our prior expectations and are headed even higher into next year.” A used graphics card is the only graphics card whose memory was bought at pre-shortage prices.

About the RTX 50 Super cards you keep reading about: NVIDIA has never announced them. VideoCardz tagged its first report (November 7, 2025) “Rumor” and noted that “NVIDIA has never confirmed a SUPER refresh”; on July 18, 2026, Tom's Hardware relayed VideoCardz's board-partner sources claiming cards had reached a partner and the launch was on hold over 3 GB GDDR7 pricing — per those sources, not NVIDIA.

The used market re-priced too: our June GPU-server guide listed the RTX 3090 at roughly $500–650; on September 4, ResalePrices put its average at $1,343 across 301 listings and GPUDojo's lowest offer at $1,570 (tagged new).

ResalePrices averages recent eBay sold and active listings (parts-only boards, bundles and outliers filtered out) — the closest thing to a transaction price; GPUDojo reports the single lowest offer across eBay, Amazon and Newegg — an asking price. Both captured September 4, 2026. First, used against the new card that replaces it, matched by PassMark's G3D Mark:

Used: ResalePrices averages and GPUDojo lowest offers, September 4, 2026; new: Tom's Hardware index, September 3; G3D Mark: PassMark, September 4. Percentages are our arithmetic.

Used at 26–65% of the new replacement; the A4000 holds 81–86%. ResalePrices, GPUDojo, Tom's Hardware, September 3–4, 2026.

The pattern: a consumer card sells used at a quarter to two-thirds of the new card that replaces it; a professional card sells at four-fifths of its own new price. An RTX 3070 delivers 77% of a new RTX 5070's G3D score for 34% of its price. The A4000 sells used at 81–86% of new, and the RTX A6000 48 GB at $3,998 used against $4,850 new (GPUDojo) is 82%. Now the ladder:

PCSP prices and units from the live catalog, re-pulled September 9, 2026 (Graphics Cards category: 57 purchasable SKUs, 55 of them GPUs). Used market: ResalePrices averages and GPUDojo lowest offers, September 4, 2026.

Ours is a workstation shelf: 24 Quadro SKUs make up 327 of the 590 units, the 294 units under $50 are 1–4 GB cards on 2010–2017 silicon (two-thirds of them Kepler and Fermi), and of 708 graphics cards listed on the site, 651 are sold out. The one card that matters to a desktop buyer is the RTX 3070: $299.99, 134 units — $31 above ResalePrices' average; the difference buys a load test, 90-day returns and same-day shipping. Speed per dollar — PassMark G3D divided by our September 5 price — tells a sharper story:

G3D Mark: PassMark, September 4, 2026 (gaming-style composite; passive cards omitted). Prices and units: PCSP catalog, September 9, 2026. Points per dollar is our arithmetic.

The top of that column is entirely legacy-branch silicon: a $90 GTX 980 Ti is twice the RTX 3070's value per PassMark point as long as nothing you run needs a driver newer than 580 or a CUDA toolkit newer than 12.x.

Cheapest per PassMark point is legacy silicon; the RTX 3070 leads mainline cards. PassMark, September 4; PCSP catalog, September 9, 2026.

Every route to 24 GB on September 4, 2026, cheapest first, with the driver status that decides bargain or paperweight in 2027:

Prices: GPUDojo (lowest offer, September 4, 2026), RigPrice (asking, checked September 5), ResalePrices (market average, September 4); new price from Tom's Hardware's index (September 3). Power figures: NVIDIA product pages and the TITAN RTX datasheet.

The spine: below the $880–895 Turing pair, every 24 GB card is either a passive data-center board or Maxwell/Pascal silicon on the last driver branch that will ever support it. The first 24 GB on mainline drivers that build with CUDA 13 is Turing — the TITAN RTX and Quadro RTX 6000 of 2018 — which is why a search for a refurbished 24 GB card lands on eight-year-old professional silicon: it is the cheapest 24 GB with a future. Above it we have nothing at 48 GB; below it, at 16 GB, the shape repeats: our only loose 16 GB cards for a PC are the passive T4 and P100, while the cheapest 16 GB on mainline drivers among the cards the trackers cover was the Turing Quadro RTX 5000 at $260 (GPUDojo).

“Is an old GPU still worth it?” has an objective answer now, because NVIDIA wrote it down. Its UNIX driver deprecation schedule, quoted by Phoronix on July 1, 2025: “The release 580 series will be the last to support GPUs based on the Maxwell, Pascal, and Volta architectures” (card list at Notebookcheck; Turing's GTX 16 stays supported). Per endoflife.date (checked September 4, 2026), R580's active support ended August 4, 2026, with security fixes to August 4, 2028; R590, the first branch without those architectures, shipped December 22, 2025. The CUDA 13.0 release notes: “Offline compilation and library support for these architectures have been removed in CUDA Toolkit 13.0,” while 12.x toolkits “will continue to be supported.” Kepler went first: support stopped after R470 (Guru3D, May 24, 2021; last builds June 4, 2024 on Linux and July 9, 2024 on Windows, per endoflife.date), and NVIDIA's legacy-GPU page now lists 470.xx for Kepler and 390.xx for Fermi. On AMD's side, Tom's Hardware reported on November 8, 2023 that regular updates for Polaris and Vega had ended that September: bug fixes continue, new features do not.

Sources: NVIDIA via Phoronix and Notebookcheck, endoflife.date, CUDA 13.0 release notes, NVIDIA legacy-GPU page, Guru3D, Tom's Hardware, Blender 5.2 manual — all linked above, checked September 4, 2026. Of our 590 units on September 9, 2026, only the Turing cards (T1000 ×16, Tesla T4 ×10, Quadro RTX 6000 ×2, four singles), the Ampere RTX 3070 ×134 and RTX A2000 ×2, and the RDNA W6600 ×10 / W6400 ×1 (plus the Apple-only W5700X ×1) are on mainline drivers.

What “legacy” means in practice:

• Display output: forever. Windows 11's graphics requirement is “compatible with DirectX 12 or later with WDDM 2.0 driver” (Microsoft), which every Kepler-or-newer NVIDIA and GCN-or-newer AMD card meets. What stops an old workstation is Microsoft's supported-processor list, not the GPU: Xeon W-2100/W-2200 and Xeon Scalable towers (Precision 5820/7820, Z4 G4, P520) are on it, with TPM 2.0 enabled; Xeon E5 v3/v4 towers (T5810, Z440, P510) are not.
• Today's creative software: fine on Pascal, not on Kepler. DaVinci Resolve 20's minimums (Blackmagic) ask for “OpenCL 1.2 or CUDA 12.8,” “NVIDIA Studio driver 570.65 or newer” and 4 GB of VRAM — a Pascal card on the 580 branch clears all three today. Blender 5.2's Cycles GPU rendering needs compute capability 5.0 (Maxwell and newer; Kepler is out), and AMD HIP “requires a AMD graphics card with the RDNA1 architecture or newer,” so a Polaris WX 7100 renders on the CPU.
• Tomorrow's CUDA software: no. Anything compiled against CUDA 13 or later has no Maxwell, Pascal or Volta target. For AI work you expect to keep updating through 2027, Turing is the floor.

Kepler ended in 2024; Maxwell/Pascal/Volta security-only after August 4, 2026. NVIDIA via Phoronix, endoflife.date, September 4, 2026.

VRAM decides most used-card purchases; the requirement is published by people with no card to sell you:

Requirements: Microsoft Windows 11 specifications; Puget Systems engineering roundup (December 18, 2025) and Resolve recommendations (March 26, 2026); Blackmagic Resolve 20 minimum requirements; Ollama library pages; Tom's Hardware best-GPU guidance — all checked September 4, 2026.

The local-model sizing is a rule of thumb from download sizes; our local LLM hardware guide does the model-by-model version. For CAD, Puget's December 2025 roundup found “minimal difference between models within a given family” because the work is CPU-bound — a $213.74 T1000 is not a compromise in a modeller's machine.

The second most common way a used-card purchase fails is a good card in a tower whose power supply has no cable for it. The OEM documents in our technical library settle this in advance (checked September 4, 2026):

Sources: Dell Precision 5820, 7820 and 7920 spec sheets and owner's manuals; HP QuickSpecs for the Z4 G4 (v4, 2018 — later revisions add RTX cards) and Z440; all checked September 4, 2026.

The card side comes from NVIDIA's product pages (board power, “required system power,” connectors; checked September 4, 2026): RTX 3070 220 W, 650 W system, one 8-pin; RTX 3090 350 W, 750 W, two 8-pins, three slots; RTX 4090 450 W, 850 W, “3x PCIe 8-pin cables” or a 450 W-plus Gen5 cable, three slots. On the professional side (NVIDIA's line-up page): the RTX A4000 is 140 W in a single slot; the RTX A2000 is 70 W, low-profile; the T1000, T600 and T400 are 50, 40 and 30 W in a single-slot low-profile card 2.713 by 6.137 inches — the usual answer to “used low profile GPU” searches.

What we can sell you to fix a power problem, as of September 9, 2026, from the power supplies category: HP Z640 925 W $29.99 (22 in stock), HP Z8 G4 1,125 W $54.99 (20), HP Z840 850 W $39.99 and $40.60 (4 across two part numbers). For the Precision 5820 we stock only the 425 W unit ($19.00 ×8, $19.94 ×4), not the 950 W; nothing for the 7820/7920 towers, the Z4 G4 750 W and 1,000 W, or the Z440 700 W. For those platforms, buy the machine with the right supply in it; empty towers are our deepest stock: a Precision 5820 barebones at $124.99 (794 units) or a 5820 with a Xeon W-2135 and 32 GB at $363.94 (235). That W-2135 5820 plus our RTX 3070 is $663.93 for a six-core, 32 GB, 8 GB-GPU tower — if the listing's supply is the 950 W unit, so ask. Configured with an 8 GB card: a Precision 5820 with a Quadro P4000 at $621.66 (94 units), or with a Xeon W-2255, 64 GB, 1 TB NVMe and a Quadro RTX 4000 at $1,777.16 (43). Every one of those carries the 1-year warranty on the card as well as the box; prices by brand and tier are in our workstation prices guide.

“Where to buy used GPUs” is really a question about what happens when the card is dead on arrival.

1. Read the return terms before the listing. On eBay, the Money Back Guarantee covers an item not as described for “30 calendar days after the estimated or actual delivery date”; the eBay Refurbished program adds an Allstate-serviced warranty (two years for “Certified,” one year for the other grades), but “the 'Seller Refurbished' item condition is not part of the official eBay Refurbished program.” Ours, verbatim from the category page: “Every card is tested under load before listing — not just a boot check — and covered by 90-day returns.” The return policy refunds shipping on a defective item; the 1-year warranty covers “all components in configured units,” and “parts purchased loose are not eligible.”

2. Ask what the card did for a living. The best public test of ex-mining cards is still Linus Tech Tips' purchase of 19 used cards, reported by Tom's Hardware on February 5, 2023: with a 3%-below-reference threshold, “nearly all of the 19 cards passed”; the failures were a Gigabyte RTX 3060 with failed VRAM and temperatures past 100 °C, and a Sapphire RX 5700 XT 3% slow in Kombustor but fine in games. Tom's caveat is ours — the sample “is pretty small” — and so is its conclusion: “make sure wherever you're getting the card from has a return policy.”

3. Look at the power connector, both ends, on any RTX 40 or 50 card. NVIDIA's statement of November 21, 2022, as reported by KitGuru, counted “50 cases globally … about 0.04% of all the RTX 4090 cards sold (130,000 units)” of melted 12VHPWR connectors, every RMA'd cable showing “a visible wear line in the connector, suggesting the cable wasn't fully inserted.” It did not end with the 4090: der8auer measured one wire on an RTX 5090 Founders Edition carrying “over 250 W of power and roughly half of the total load” (PC Gamer, February 11, 2025), and Club386 destroyed its own RTX 5090 FE and a 1,000 W supply with a dedicated single cable (June 22, 2026). A used buyer inherits someone else's insertion habits: demand a photo of the card's socket under light and look for discolouration or that wear line; no photo, no sale. RTX 30-series (bar the 16-pin RTX 3090 Ti), Quadro RTX and RTX A-series cards use the older 6-, 8- and 12-pin connectors, not 12VHPWR.

4. Test it in the first week, not the last. The return window is the warranty. Sensor sanity first, with GPU-Z or HWiNFO: BIOS and memory vendor match the listing, fan speed rises with load, the edge-to-hot-spot temperature gap is sane. Then a 15–30 minute stress run — FurMark 2 or OCCT v17, whose 3D adaptive test is built for “picking up any artefact or errors during execution” — with clocks steady, no artefacts and temperatures inside NVIDIA's figures for the card. Finish with a VRAM pass (OCCT's memory test or memtest_vulkan): the LTT failure above was memory, and memory faults hide behind a clean boot.

5. Match the physical card to the listing. Part number on the bracket against the OEM number in the title (our RTX 3070 is Dell part DD32D); outputs against what you need; bracket height, length and, for the RTX 3090 and 4080/4090 class, three-slot thickness against the chassis.

Half of our shelf is Quadro; we have a stake in this. NVIDIA's ISV certification page defines “Certified” as application, driver and hardware combinations “formally tested against criteria set forth by ISV application developers,” with the consequence that “end users have access to direct support from the application developer (and NVIDIA, if needed).” Two hardware differences ride along: ECC (NVIDIA lists the RTX A4000 with “16GB GDDR6 with error-correction code”; no GeForce page lists it) and the single-slot and low-profile form factors above, which no GeForce of comparable class offers.

Against that, the testers. Puget Systems' DaVinci Resolve recommendations (March 26, 2026): “for most users a GeForce card is the better option. Not only are GeForce cards much more affordable, they are able to match or beat the performance of the pro-grade cards,” the pro cards' advantage being “much higher VRAM capacities.”

Our rule when asked: buy the used professional card for certification, ECC, single-slot or low-profile fit, or four-plus display outputs; buy the used GeForce for speed per dollar; and expect the professional card to hold its value. One name comes up constantly: the Quadro T2000 is a laptop GPU with no desktop version; the desktop equivalents are the T1000, T600 and T400.

We sell used cards; read this as the disclosure it is.

• 1440p and 4K gaming, or ray tracing without compromises. Tom's Hardware wants 12–16 GB for that; every GeForce we stock has 8 GB or less, bar one Pascal TITAN Xp 12 GB on the legacy branch. A new Radeon RX 9060 XT 16 GB at $469 (Tom's index, September 3), with a warranty and an invoice, is the better buy than any used 8 GB card — ours included.
• Software that will require CUDA 13 or newer. No Maxwell, Pascal or Volta card, however cheap, however much VRAM: not the $128.65 P100, not the $239 P40, not a P6000. If the toolchain is moving, the floor is Turing.
• Anything above 8 GB in a desktop, today, from us. Between $299.99 and $1,092.49 we hold one active card, a 6 GB RTX A2000 at $348.99 (two units); nothing at 48 GB. For an RTX 3080, 3090, 4070, 4090, A4000, A5000 or A6000, the used market has them at the prices in this article and we do not.
• Used RTX 40 and 50 cards you cannot inspect. The 12VHPWR history above is a history of insertion errors, and a used card comes with the previous owner's. Buy with a return window and a photo, or buy new.
• Passive Tesla cards in a desktop. The P100, T4, M10 and M60 on our shelf were designed for a server's front-to-back airflow; in a tower without a ducted fan they throttle or die. If you do not already know why you want one, you do not want one.
• Small-form-factor builds. Our low-profile stock is four SKUs and eight units; our two RTX A2000 6 GB ($348.99) ship with the full-height bracket, and the 12 GB A2000 we do not have.
• A card you need under warranty for a year. Loose parts get 90-day returns, not the 1-year warranty; if the year matters, buy the card inside a configured workstation, or buy new.
• AI that runs a few hours a month. Renting beats owning below a utilization threshold — the break-even arithmetic is in our AI server for business guide — and a used 24 GB card at $880–1,343 is on the wrong side of it.

Are used graphics cards good?

The best public evidence says yes, with a return window. Linus Tech Tips bought 19 used cards, including ex-mining RTX 3060/3070/3080 boards, and “nearly all” passed a 3%-below-reference test; the hard failure was an RTX 3060 with overheating VRAM (Tom's Hardware, February 5, 2023). Memory is the usual failure point, so test VRAM inside the return period.

Where is the best place to buy a used GPU?

Wherever the return terms are written down. eBay's Money Back Guarantee gives 30 days for an item not as described; its official Refurbished program adds a one- or two-year Allstate-serviced warranty, but “Seller Refurbished” listings are outside it. Trackers such as ResalePrices and GPUDojo show the going rate. Our cards are tested under load with 90-day returns; inside a configured workstation, the card gets the 1-year warranty.

How long do used GPUs last?

Physically, well beyond their software life; the limit is driver support. NVIDIA's 580 branch is the last for Maxwell, Pascal and Volta (2014–2018), with security fixes to August 2028 per endoflife.date, and CUDA 13 no longer compiles for them; Kepler stopped getting updates in mid-2024. Turing and newer stay on mainline drivers. Buy Turing or newer for a card with a future; Pascal only for what it runs today.

Is a Quadro T1000 good for CAD or office work?

For CAD modelling, yes: Puget Systems' 2025 roundup found model creation in Revit, Inventor and SOLIDWORKS CPU-bound, and the T1000 is a 50 W single-slot, low-profile Turing card on mainline, ISV-certified drivers. It is not a gaming or rendering card (PassMark G3D 7,659 against 22,077 for an RTX 3070). Ours was $213.74 with 15 in stock on September 9, 2026.

Where can I find a refurbished 24 GB GPU?

On September 4, 2026 the used market offered 24 GB in three bands: passive or legacy-driver cards under $600 (Tesla M40 $229, Tesla P40 $239); Turing cards on mainline drivers at $880–950 (TITAN RTX, Quadro RTX 6000 — the latter $1,092.49 on our shelf, two in stock); and Ampere/Ada from $1,343 (RTX 3090) to $2,500 (RTX 4090). A used RTX A6000 48 GB was $3,998 on eBay.

Is an old GPU still worth it in 2026?

For display output and today's software, often yes; for anything built on new CUDA, no. A Pascal card such as the Quadro P2200 ($99.99) still runs DaVinci Resolve 20 and Blender 5.2 Cycles, but it is absent from NVIDIA's 590-and-later drivers and from CUDA 13 builds, so its software life is what exists now. Kepler and Fermi cards (Quadro K600, Quadro 600, GTX 770) are display adapters in 2026: legacy drivers, no current CUDA, no Blender GPU rendering.

What should I check when buying a used graphics card?

Five things: the seller's written return terms; what the card did (ex-mining is not disqualifying, but the LTT test's hard failure was VRAM); the power connector on any RTX 40/50 card for discolouration or the wear line NVIDIA described; a 15–30 minute FurMark 2 or OCCT run with steady clocks and no artefacts, then a VRAM test; and the physical match — part number, outputs, size, power cable.

Do used graphics cards come with a warranty?

From most sellers, no — a return window is the norm; eBay's Refurbished program is the exception, with a one- or two-year warranty on graded cards. At PCSP a loose card gets 90-day returns (defects refunded with shipping; a working card returned in error carries a 20% restocking fee), not the 1-year warranty; a card bought inside a configured workstation, desktop or server is covered for the year.