TL;DR: A used RTX 4090 averages $2,268 in August 2026 — 79% more than a used RTX 3090 at $1,264, for the same 24GB of VRAM and only 8% more memory bandwidth. For LLM chat and coding, that premium buys you roughly 10–40% more tokens per second. It only makes clear sense if image generation or fine-tuning is your main workload.

	Used RTX 4090 ($2,268)	Used RTX 3090 ($1,264)	RTX 5090 ($4,699)
Best for	Diffusion + fine-tuning + LLM in one box	Pure LLM inference on a budget	Max single-card speed and 32GB
VRAM / bandwidth	24GB / 1,008 GB/s	24GB / 936 GB/s	32GB / 1,792 GB/s
The catch	79% price premium for an 8% bandwidth bump	~2× slower diffusion, no FP8	Costs more than two used 4090s’ worth of rig

Honest take: If your GPU exists to serve tokens from a quantized model, buy the 3090 and pocket the $1,000 — decode speed follows memory bandwidth, and the 4090 barely has more. If you run FLUX or SDXL daily or fine-tune with QLoRA, the 4090’s compute is real and worth paying for. Almost nobody should pay $3,400+ for remaining “new” 4090 stock.

A four-year-old GPU that NVIDIA stopped making now sells used for 42% more than it cost new. The RTX 4090 launched at $1,599 in October 2022, was discontinued when the RTX 40 series wound down in late 2024, and in August 2026 averages $2,268 on the used market. That price is not nostalgia. It’s 24GB of VRAM in a market where the memory crunch has pushed every alternative up even harder. The question is whether you should pay it — and for most local-LLM workloads, the answer is no.

The August 2026 price board

Three cards define the 24GB-and-up decision right now. Prices below are US used-market averages as of August 2026:

Card	Used price (Aug 2026)	VRAM	Bandwidth	$ per GB VRAM	GB/s per $
RTX 3090	$1,264 avg, $1,201–$1,299 fair range, 366 listings	24GB	936 GB/s	$52.70	0.74
RTX 4090	$2,268 avg; fair asking $2,200–$2,500	24GB	1,008 GB/s	$94.50	0.44
RTX 5090	$4,699 median street, up from $4,300 in June	32GB	1,792 GB/s	$146.90	0.38

The dollar-per-gigabyte and bandwidth-per-dollar columns are our arithmetic from those tracked prices, and they tell the whole value story: the 3090 delivers 68% more bandwidth per dollar than the 4090. The used 3090 market is also up 6.6% over 90 days, so neither card is getting cheaper while you wait.

One trap to skip entirely: leftover “new” 4090 stock. With production ended, the few sealed cards still listed at authorized resellers run $3,400 and up — channel remainders, not fresh silicon. At that price you’re $1,300 from a 5090 with 32GB. Either buy used or change tiers.

We covered why the whole market moved in our H200/China supply analysis — the short version is that the memory supercycle has no 2026 relief in sight, which is why “wait for prices to drop” hasn’t worked all year.

Decode speed: the 8% problem

Here’s the fact that should anchor the whole decision: LLM decode — the tokens-streaming-at-you part — is memory-bandwidth-bound, not compute-bound. Every generated token requires reading the active weights from VRAM. The 4090 reads at 1,008 GB/s, the RTX 3090 at 936 GB/s. That’s a 7.7% difference on the spec sheet, and single-stream chat speed tracks it far more closely than the price gap suggests.

Measured numbers bear this out, with spread. Hardware Corner’s RTX 4090 test bench (llama.cpp, Ubuntu 24.04, CUDA 12.8) puts Qwen3 8B Q4 at 104 tok/s with a 16K context on the 4090. Cross-source roundups that test both cards at short context report the 4090 around 135 tok/s on 7B Q4 against the 3090’s ~95 tok/s. Published 8B-class figures for the 4090 cluster anywhere from 95 to 145 tok/s because context length, quant, and runtime version all move the number.

The honest range, then: expect the 4090 to be roughly 10–40% faster at interactive LLM work, with the small end of that range showing up in exactly the scenario most people buy for — a single user, a quantized model, a long context. At reading speed (~7–10 tok/s), both cards produced text faster than you can consume it years ago. You are paying $1,004 extra for a speed difference you’ll mostly notice in benchmarks.

Both cards run the same models, because what fits in 24GB is identical: Qwen3.6 35B-A3B, Gemma 4 26B QAT, gpt-oss-20b, 27B-class dense models at Q4 with room for context. The 4090 changes none of that list.

Where the 4090 is genuinely 1.5–2× better

The premium stops being irrational the moment your workload becomes compute-bound, because Ada’s tensor throughput is roughly 1.5–2× Ampere’s. Three workloads live there:

Image and video generation. Diffusion is compute-heavy, and it shows: across 12 aggregated Stable Diffusion/SDXL/FLUX benchmarks, the 4090 is a median 46% faster than the 3090, with individual tests running 40–70% depending on sampler and resolution. If ComfyUI is open on your machine more than your chat UI, the 4090’s premium roughly matches its output advantage.

Fine-tuning. Training passes are compute-bound too. When we ran the total-cost math on QLoRA fine-tuning, the 4090’s step-time advantage compounded across every epoch — the difference between an overnight run finishing at 3 a.m. versus 7 a.m.

FP8. Ada’s fourth-generation Tensor Cores add hardware FP8, which Ampere simply lacks. TensorRT-LLM and vLLM FP8 quants use it; on a 3090 those paths fall back or fail. This matters more every quarter as FP8 checkpoints become the default fast path for new releases — part of why local models got so much faster in 2026.

If none of those three paragraphs describes your week, you’ve just read $1,004 of features you won’t use.

The capacity play: two 3090s beat one 4090

Same money, different shape: two used 3090s cost about $2,528 — $260 more than one used 4090 — and give you 48GB of VRAM instead of 24GB. That’s the difference between offloading a 70B model painfully and holding a 70B Q4 (~42.5GB) fully resident. The 3090 also supports NVLink, which the 4090 dropped entirely — dual-4090 setups are PCIe-only.

The catch is that dual-GPU is a project, not a purchase: BIOS PCIe settings, IOMMU/ACS behavior, and slot topology can silently halve your gains, and llama.cpp layer-split on dual 3090s lands around 7–10 tok/s on 70B Q4 in the community reports we verified for that guide. But if capacity is the goal, no single 24GB card at any price competes with $2,528 of paired 3090s.

Buying used: the one component you must inspect

Every used 4090 carries a known hardware risk the 3090 doesn’t: the 12VHPWR power connector. NVIDIA acknowledged around 50 melted-connector cases early on, and reports continued for years — including a reviewer’s card that ran fine for two years while the connector had quietly melted. The failure mode is an incompletely seated connector heating up under the card’s 450W draw.

Before money changes hands, check three things:

1. The socket itself. Pull the cable and look into the 16-pin socket on the card for discoloration, warped plastic, or burn marks. A melted connector can hide behind a working card.
2. Load behavior. Ask for (or run) a sustained load test. Random black screens under load are the classic symptom.
3. Actual power draw. Under inference load, verify the card pulls what it should:

$ nvidia-smi --query-gpu=power.draw,power.limit,temperature.gpu --format=csv
power.draw [W], power.limit [W], temperature.gpu
382.45 W, 450.00 W, 67

A 4090 that tops out far below ~350–400W under a real inference load, or that thermal-throttles immediately, has a story the seller isn’t telling. Measured draw during LLM work runs 360–410W against the 450W TDP.

If the card uses the revised 12V-2x6 connector (later production) rather than original 12VHPWR, the risk profile improves — worth asking the seller which it has.

Running cost: the difference is smaller than the sticker

At the August 2026 US average residential rate of 18.44¢/kWh, a 4090 averaging ~385W through two hours of heavy inference a day costs about $4.26/month in electricity (0.385 kW × 60 hours × $0.1844 — our arithmetic). A 3090 at ~350W is about $3.87. Power cost won’t decide this purchase; we’ve run the full 24/7 server math elsewhere if your box never sleeps. Both cards also power-limit gracefully — a 4090 capped at 300W gives up only a small slice of inference speed, because decode wasn’t using the full 450W anyway.

Buy, skip, or rent: the actual decision

Buy the used 4090 ($2,268) if you generate images or video most days, you fine-tune locally, or you want FP8 support without paying RTX 5090 money. One card, every workload, no dual-GPU headaches.

Buy the used 3090 ($1,264) if the GPU’s job is serving tokens — chat, coding assistance via a local backend (our sister site covers wiring local models into coding tools), RAG, agents. It remains the value king we called it in May, and the $1,004 you save is most of a second 3090.

Buy two 3090s (~$2,528) if you want 70B-class models resident in VRAM and you’re willing to do the dual-GPU homework.

Buy the 5090 ($4,699) if you need 32GB on one card and maximum single-stream speed — we ran that comparison against the 4090 when the gap was $400; at today’s gap of $2,431, it’s a much harder sell.

Rent first if your heavy workloads are occasional. An RTX 4090-class or bigger card on RunPod for a weekend of fine-tuning or a batch of renders costs a few dollars and tells you whether the compute-bound story actually describes your usage before you commit $2,268 to it. Our rent-vs-buy framework puts numbers on the crossover.

FAQ

Is a used RTX 4090 worth it for LLM inference alone? Generally no. Decode speed follows memory bandwidth, and the 4090 has only 7.7% more than a 3090 (1,008 vs 936 GB/s) at 79% higher cost. Real-world interactive speedups run roughly 10–40% depending on model and context. The 3090 is the better pure-inference buy at August 2026 prices.

Why is the RTX 4090 more expensive used in 2026 than it was new in 2022? Production ended in late 2024, 24GB consumer cards became the local-AI workhorse tier, and the 2026 memory-price crunch lifted every VRAM-heavy card. $1,599 MSRP became a $2,268 used average with no new supply arriving.

Should I worry about the melting connector on a used 4090? Take it seriously but don’t let it kill the deal: inspect the 16-pin socket for discoloration or warping, run a sustained load test, and confirm 360–410W draw under load. Cards with the revised 12V-2x6 connector carry less risk.

Does the 4090 run bigger models than the 3090? No. Both have 24GB, so the runnable-model list is identical. The 4090 runs the same models somewhat faster, and compute-bound work (diffusion, fine-tuning) substantially faster. For bigger models, you need more VRAM — two 3090s, a 5090, or unified-memory machines.

Sources

• RTX 4090 Price Tracker US, Aug 2026 — Best Value GPU
• RTX 4090 Used GPU Price & Fair Asking Range — ResalePrices
• RTX 3090 Price Tracker US, Aug 2026 — Best Value GPU
• RTX 3090 Used GPU Price & Fair Asking Range — ResalePrices
• RTX 5090 Price Tracker, August 2026 — videocardprices.com
• Nvidia’s RTX 40-series is coming to an end — Digital Trends
• RTX 4090 Local LLM Benchmarks, Context Scaling & Supported Models — Hardware Corner
• LLM Tokens/Sec Benchmarks: RTX 4090 vs 3090, 7B–70B Q4 (llama.cpp) — Mustafa.net
• RTX 3090 vs RTX 4090 for AI: Performance, FP16 Throughput & Upgrade Analysis — BestGPUsForAI
• GPU Power During AI Inference by Model — GIGAGPU
• Electricity Rates by State: August 2026 Report — ChooseEnergy (EIA data)
• NVIDIA: ~50 cases of melted 12VHPWR connectors on RTX 4090 — TweakTown
• Official RTX 4090 power cable found melted by reviewer 2 years later — Tom’s Hardware
• RTX 4090 for AI/ML: Benchmarks, Specs, and Pricing — Spheron

Last updated August 23, 2026. GPU prices move weekly — verify current listings before purchasing.

Recommended Gear

• RTX 4090 — the one-card-does-everything pick, if the used price and connector check out
• RTX 3090 — still the bandwidth-per-dollar king for pure LLM inference
• RTX 5090 — 32GB and 1,792 GB/s, for when the model won’t fit anything else