Field notes / № 3

Field note № 3 · Verified 2026-07-05

Does a used RTX 3090 still pay for itself in 2026?

The RTX 3090 launched in 2020. Six years later it is still the default answer on every local-AI forum, because nothing else puts 24 GB of VRAM in your hands for about $750. We rebuilt the whole rig on paper, metered it honestly, and ran it through our break-even engine against six real cloud baselines. Same rig, same electricity — three different verdicts. Here they are.

Why the 3090 stayed the people's GPU

VRAM is the whole game in local inference. A model either fits in memory or it doesn't; everything else is negotiation. NVIDIA has sold exactly one consumer card that combined 24 GB with a used price ordinary people will pay, and it's this one. The 4090 has the same 24 GB for roughly double. The 5090 has 32 GB for roughly triple. Meanwhile a used 3090 sits at about $750 on the used market (verified 2026-07-05 — see d-central's used-3090 analysis and compute-market's 2026 GPU pricing).

It's not fast by 2026 standards. It is sufficient, and sufficiency at $750 beats excellence at $2,400 for most people. That's the entire thesis. Now let's check whether the math still agrees.

The full build: about $1,200

A GPU is not a computer. Here is the complete parts list from our used-3090 preset, priced at street rates on 2026-07-05:

The engine rounds this to $1,200; real-world builds land anywhere from $1,100 to $1,300 depending on how patient you are with listings. Don't cheap out on the PSU — a 3090's transient spikes eat budget units for breakfast.

What it actually runs

24 GB comfortably holds models up to about 32B parameters at Q4 quantization. 70B is technically possible with aggressive quantization and CPU offload, but the speed makes it a demo, not a tool. Measured throughput for the preset:

A well-tuned 32B in 2026 is genuinely good — a fair sparring partner for mid-tier cloud models on coding and everyday reasoning. It is not a frontier model, and we won't pretend otherwise. If your work truly needs frontier quality, the honest comparison changes — more on that in the six-baseline table below.

What owning it really costs: $42 a month

Our methodology counts two things: hardware amortization and electricity at the wall, idle hours included. For the coding-assistant usage preset — 3 hours of inference and 5 powered-but-idle hours a day, about 15M tokens a month — the engine says:

• Amortization: $1,200 over 3 years = $33.33/mo. The rig isn't free just because you already paid for it.
• Electricity: 450 W active, 50 W idle → 48 kWh/mo. At the EIA national average of $0.1883/kWh, that's $9.04/mo.

Total cost of ownership: $42.37 a month, flat — no matter which model you run on it. That number is the whole left side of every comparison that follows.

Your state moves the electricity line more than your GPU choice does — we did the full electricity teardown in Field note № 1.

Break-even against six clouds

Here's where every affiliate calculator quietly cheats: they compare your rig against the most expensive API they can find. We compare against the model you'd actually use. Same $1,200 rig, same coding workload (7.5M input + 7.5M output tokens/mo), national electricity — six honest baselines, engine-computed:

Read the last row twice. DeepSeek V4 Flash costs $3.15 a month at this workload — less than the rig's electricity bill. Against that baseline the 3090 never pays for itself. Not slowly. Never. If a cheap open-weight API already covers your needs, we'd rather tell you to keep renting than sell you a graphics card. That said — one caveat in the rig's favor: a 32B local model isn't competing with frontier quality, so the fairest fights on this table are the Sonnet, GPT-5.4, and Haiku rows. Against the mid-tier, the used 3090 wins in under ten months. We unpacked why the baseline matters this much in Field note № 2.

Usage moves the verdict too

The other lever nobody shows you: how much you actually use the thing. Same rig, same national rate, four usage presets from our engine, against two baselines:

The pattern is blunt. A casual chatter should not buy this rig — $21/mo of Sonnet usage takes 68 months to claw back $1,200, and the card will be nine years old before it breaks even. A daily coder breaks even inside a year. An always-on agent server breaks even before the thermal paste settles. Hardware rewards heavy users; light users are exactly who the buy-a-GPU sites shouldn't be talking to, and exactly who they target.

The 70B itch: dual 3090s

If 32B isn't enough, the classic move is a second card: two used 3090s, 48 GB pooled, about $2,250 all-in, running 70B at Q4 at a usable 15 tok/s. The engine's verdict at the coding workload: against Claude Opus 4.8 it breaks even in 10.7 months — OWN. Against Sonnet it slips to 18.8 months — HYBRID territory. Push it to always-on agent duty and it pays for itself against Sonnet in 3.7 months.

Two honest warnings. Dual-GPU setups are the hardest build we model (setup score 3/10 — budget a weekend and some forum posts), and 750 W of active draw is a space heater with a hobby. It's the right call for heavy users who specifically need 70B-class quality. It's the wrong first rig.

The honest downsides

The break-even math above is real, but it isn't the whole decision. Before you open eBay:

• Used-market roulette. Many 3090s mined crypto for years. GDDR6X runs hot and the factory thermal pads age badly. Buy from sellers with return windows, stress-test in week one, and budget $30 for replacement pads.
• No warranty. When a $750 used card dies at month 14, the math you just read changes retroactively. That risk isn't in the engine because it can't honestly be averaged.
• Power and heat. A 350 W GPU means a ~450 W system under load, with spikes beyond. Your office gets warmer. In summer, your AC pays part of the bill the engine doesn't see.
• Setup effort. This is a 6/10-ease build: fine if you've assembled a PC before, a real weekend if you haven't. Quantization settings and driver versions are part of the purchase price; they're just paid in patience.
• Cloud prices keep falling. Every API price cut extends your break-even. The trend has one direction, and it isn't ours.

None of these show up in the dollar math, and none of them should be hidden behind it. There's also the other side of the ledger: local means your prompts never leave the machine, no rate limits, and a model nobody can deprecate on a Tuesday. We score that separately — sovereignty never gets smuggled into the dollars.

The verdict

So: does a used RTX 3090 still pay for itself in 2026? Yes — if you're a heavy user replacing a mid-tier or frontier API. No, if you chat casually or would otherwise pay $3 a month for an open-weight endpoint. The 3090 remains the best value in local AI precisely because the question has a real answer either way, and most sites only tell you half of it.

Your electricity rate, your usage, and your real cloud baseline will move every number on this page. So run yours:

Run your own numbers

Or start from a worked example: used 3090 vs Claude Sonnet in Texas — then swap in your own state and baseline.

All prices verified 2026-07-05: used-GPU street prices, EIA residential electricity rates, and published API list prices. Every dollar figure above comes from the same open engine that powers the test — the formulas are in our methodology, and the caveats are in our honesty policy. Prices drift; check live links before you buy anything.