ToolScore
Local GPU AI 5 min read

GPU Rent vs Buy: The 5,700-Hour Math on a Used RTX 3090

~5,700 hours to break even the number that matters here

By Stan K.

Renting a 3090 runs about $0.21/hour (vast.ai median) while used cards sell at $1,199–$6,919. The break-even math with electricity included for the 3090, 4090 and 5090 — and why the answer is always a hybrid.

A used RTX 3090 sells for about $1,199 (sold median), and renting one runs about $0.21/hour on vast.ai — so buying only pays off after ~5,700 GPU-hours, before you have paid a single cent of electricity. At four hours a day, that is roughly four years. Rent versus buy is not an ideology; it is arithmetic with three inputs and a mortality risk.

The three inputs

Rental rates (on-demand, checked 2026-11-10). Vast.ai medians: RTX 3090 ~$0.21/hour, RTX 4090 ~$0.48/hour, RTX 5090 ~$0.65/hour. Individual listings swing around those medians — the vast floor dips to $0.122/hour for a 3090 and $0.213/hour for a 5090, and peak-demand rates run higher.

Used purchase prices (US sold medians, human-verified 2026-11-10). RTX 3090: $1,199 ($908–1,343) — the 24GB AI premium has inflated this card hard (the marketplace recon traces the same card at $700–900 not long ago). RTX 4090: $2,387 ($2,144–3,000). RTX 5090: essentially no used market — new retail runs $6,549–9,490, with the scalped street median at $6,919.

Electricity — the cost renters never see. Inference pulls roughly rated board power for as long as it runs: ~350W (3090), ~450W (4090), ~575W (5090). At the US average of ~$0.18/kWh and 4 hours/day of actual compute:

  • 3090: 511 kWh/year ≈ $92/year
  • 4090: 657 kWh/year ≈ $118/year
  • 5090: 840 kWh/year ≈ $151/year

A box that runs 24/7 multiplies those by six — about $552/year for a 3090. Rental prices have the host’s electricity baked in; ownership bills you separately, forever.

The math, concretely

Break-even hours = used price ÷ hourly rent. At vast.ai median rates:

  • RTX 3090: $1,199 ÷ $0.21 ≈ 5,700 hours
  • RTX 4090: $2,387 ÷ $0.48 ≈ 5,000 hours
  • RTX 5090: $6,919 ÷ $0.65 ≈ 10,600 hours
Rental hours that equal the purchase price (vast.ai median rates)
RTX 3090 — $1,199 / $0.21/h 5,700 h
RTX 4090 — $2,387 / $0.48/h 5,000 h
RTX 5090 — $6,919 / $0.65/h 10,600 h

Electricity pushes every one of those numbers further out: the ~$0.06/hour a 3090 costs to feed at US rates is real money the renter’s hourly price already includes.

~5,700 hto break even on a used RTX 3090

$1,199 ÷ $0.21/hour (vast.ai median). That is ~4 years at 4 hours/day, or ~8 months at 24/7 — before electricity, before failures, before the hours you billed yourself for babysitting drivers.

The risks on both sides

Buying is a bet against hardware mortality. The cautionary tale of the season is the CMP 170HX: a community operator reported 2 of 5 cards dead within two weeks, a third arriving with defective tensor cores. Exotic silicon carries exotic death rates — and consumer cards have their own used-market traps: mining history, 3090s overdue for a repaste, “for parts” listings dressed as deals. Demand a video test before paying. Then add electricity and your own time keeping the box alive.

Renting is a bet you can stop. You pay for idle: setup, image pulls, forgotten pods, plus the quiet add-ons — persistent volumes, egress, snapshot storage. And past ~5,700 hours you have fully bought a 3090. For someone else.

The rules the forums already derived

The heavy users have done this math on real fleets, and their rules compress well:

  • Under ~40% sustained utilization → rent. Over ~60% → buy (two-year horizon, worked out on H200 nodes).
  • Under ~1B tokens/month → neither: API pricing wins outright.

The hybrid is the answer

The break-even table is why “which one?” is the wrong question. Own the small, always-on tier: one or two RTX 3060 12GB ($319 sold median, $210–250 locally each — see the budget build under $800 and the 2x 3060 vs 3090 comparison) cover learning, TTS and image generation with privacy and zero metering — things renting and APIs structurally cannot give you. Rent the 24GB+ tier by the hour for spikes (the over-$800 used-GPU guide maps that tier). Route heavy daily volume to an API or a coding plan — that trade-off gets its own article in rented GPU vs API.

The takeaway

A rule that survives price swings: buy the machine you will run every day, rent the machine you will run twice. The 3090 at $1,199 was a $700–900 card recently and may be again; the ~5,700-hour break-even just moves with it. What does not change is the structure: ownership prices your discipline, rental prices your spikes, and the hybrid keeps both bills small.

Rental rates are vast.ai median on-demand prices; used prices are US sold medians, human-verified 2026-11-10 — online ask/BIN runs higher. Electricity at ~$0.18/kWh US average; substitute your local rate.

Numbers change — the math doesn't
Token prices and plan limits shift monthly; the structural arguments in this guide age slowly. For current per-tool numbers see the live rankings.