September 2026 GPU Price/Performance Rankings: Resale Is Now Cheaper Than Retail on Every RTX 50 Card

The Advice Just Inverted: Resale Beats Retail

I normally end the RTX 50 section with a warning to go check Microcenter or Newegg before you pay an eBay seller's premium. That warning is now reversed. Here is August best-deal pricing next to what the same card cost new on Amazon in August, both from our own data:

Resale wins on all seven. On the 5090 it wins by $1,008, which is 24% off the price of buying the same card new. Both columns are August, both use the same method (the average of the 3 cheapest listings), so this is a like-for-like comparison rather than one of those apples-to-oranges price screenshots. The one thing I would not overclaim: I can tell an Amazon listing is new, but I cannot always tell whether the seller is Amazon itself or a third party on their marketplace, so read that column as "what a new one cost on Amazon" rather than as a manufacturer's storefront price. For what it is worth, an independent price tracker had these same cards within a few percent of my numbers.¹ The cause is not scalping this time, it is the memory market: AI datacenter demand has made GDDR7 scarce enough that Nvidia is reportedly allocating it by revenue per gigabyte, with no new consumer generation expected until late 2027 at the earliest.¹ New stock repriced upward to match, and the used market, full of cards bought at 2025 prices, has not caught up yet.

Measured against MSRP the chart still tells a grim story: the 5090 sits 63% over sticker, the 5060 Ti 18% over, and the 5070 Ti 13% over. The chart shows the six largest premiums, which means the one card that came in under MSRP does not appear on it: the 5070 at $539, 2% below its $549 sticker, and that is the entire below-MSRP list. Honestly, MSRP is becoming a historical footnote for this generation rather than a price anyone can pay.

1440p Gaming Best Bang for Your Buck in September 2026

The cheapest route to good 1440p frames is the RTX 3060 Ti at $1.05/FPS, and it is cheaper than last month. Of course the catch is memory. At only 8GB, newer releases slow if you turn textures up. For about $140 more, the RX 6800 XT gets you to 16GB, and the RX 7800 XT is the faster of the two if you can stretch to $412. Both cost more per frame, but if you expect to keep the card a couple of years, that is where I would put the money.

4K Gaming Best Bang for Your Buck in September 2026

The top of this chart is a trap. The RTX 3070 Ti wins on cost per frame at $2.40, but it only wins because Counter-Strike 2 is light on memory where modern AAA 4K games with textures up are not. I would take the RX 7900 XT at $2.47 instead: near enough the same cost per frame, 20GB behind it, and 38% below its original MSRP. If $558 is too much, the RX 7800 XT is the cheapest card here with real memory behind it.

AI Inference Best Bang for Your Buck in September 2026

Intel leads on paper. The Arc B580 tops the chart at $1.03 per INT8 TOP, and the catch is not the hardware. OneAPI and IPEX-LLM work, but far less of the open source AI stack assumes them than assumes CUDA, and you will feel that on the day something does not build. The safe path got cheaper this month: the RTX 3070 fell 16% to $1.09/TOP, the cheapest CUDA throughput on the board. Both are 8 to 12GB cards though, and the cheapest 16GB option is the RX 9060 XT at $1.84. Once a model does not fit, the rate stops mattering.

LLM Training and Fine-Tuning Best Bang for Your Buck in September 2026

Ignore the winner. The Tesla P100 takes the chart at $6.1/TFLOP on real listings I checked, but it is a 2016 datacenter card with no tensor cores and thinning framework support. Buy it as a cheap experiment, not as a training rig. The practical leader is the RTX 4080 at $8.0/TFLOP even after rising 7% in August, and Blackwell charges 15 to 20% more per TFLOP if you want FP4. All of them are 16GB. If capacity is the real constraint, the NVIDIA A30 is the outlier worth knowing: 24GB of HBM2 at $9.8/TFLOP, roughly what a 5080 costs.

New Silicon Up, Used RTX 30 Series Down

Sort August by architecture and the split is clean. All twelve current-gen cards I track were flat or up, and most of the RTX 40 series followed. The used RTX 3070 went the other way, down 16%, and the RTX 3090 with it. It is not a universal rule, since the RTX 3080 Ti rose 10%, but the mechanism is worth understanding: the memory crunch reprices anything still being manufactured, while five-year-old cards are priced by how many people happen to be selling them. Only one of those two things is on your side right now.

Other Notes

• The used RTX 30 and RX 6000 series are still the deepest discounts. Against original MSRP the RTX 3070 is 65% below at $177, the RTX 3080 Ti is 64% below at $432, and the RX 6950 XT is 62% below at $415. Depth of discount is not the same as direction, though: only the 3070 is still falling. The 3080 Ti rose 10% and the 6950 XT rose 3% in August, so a big number next to MSRP does not mean the card is getting cheaper.
• Datacenter: the H200 is undercutting the H100. The H200 NVL best deal fell 18% to about $23.2K while the H100 PCIe eased 6% to about $26.0K, so the newer card is now the cheaper one. Take it with salt: we saw 31 distinct H200 listings and only 10 H100 listings that month, and at those volumes a couple of sellers move the whole number. Going the other way, the A100 PCIe rose 12% to $12.6K and the A10 rose 20% to $2.8K.
• Still gathering input on a memory-aware ranking. Last month I asked what VRAM thresholds and workloads matter to you, because $/TOP and $/TFLOP treat an 8GB card and a 24GB card as interchangeable. The 3070 topping the CUDA inference value list on 8GB is exactly the case where that breaks down. If you buy GPUs for AI, tell me what you need on the contact page.