Our renting-vs-buying piece covered the enterprise version of this question — $25K accelerators, ops staff, utilization curves. But the most common version of the argument now happens at a much smaller scale: someone with a desk, a game PC and a growing suspicion that they're about to spend real money on API bills or cloud GPU hours. This is the home version of the spreadsheet, run honestly.
The three lines that matter
1. The box
A used RTX 3090 with 24 GB of VRAM trades hands for roughly $700–$900 as of October 2026 (verify against your local market — prices move). You supply the rest of the machine, which you probably own. Call the incremental capex $820 to be concrete; every conclusion below scales with your real number.
2. The electricity
A 3090 draws ~350 W under load and a typical host PC idles around 40–60 W. At 8 hours of load a day plus idle the rest, you're looking at roughly 90–100 kWh a month. That's ~$15/month at the US average of $0.14–$0.18/kWh — and ~$40/month at Germany's ~$0.42/kWh. If you skipped the geography lesson: the same home lab is a better deal in Vancouver than in Berlin, through no fault of your own.
3. The rental alternative
24 GB-class consumer GPUs rent on open marketplaces for roughly $0.30–$0.50/hour, cheaper than hyperscaler consumer tiers and instantly available. That's the number your home GPU has to beat, hour by hour, including the hours the card sits idle doing nothing for you.
The break-even, roughly
| Your daily usage | Renting (24 GB class) | Owning (3090, US power) | Verdict | |
|---|---|---|---|---|
| 2 hours/day | ~$18–30/mo | ~$820 + ~$10/mo | Rent. Payback would take 2+ years and outlive your patience. | |
| 4 hours/day | ~$36–60/mo | ~$820 + ~$11/mo | Borderline. Only if you value latency, privacy or offline use. | |
| 8 hours/day | ~$72–120/mo | ~$820 + ~$15/mo | Own. Payback in roughly 9–14 months, then it's nearly free hours. | |
| Near-continuous | $220–360/mo | ~$820 + ~$40/mo | Own, obviously — and consider a second card before you consider the cloud. |
Two adjustments, both in the used card's favor. First, resale: 3090s have held value unusually well because 24 GB at 350 W keeps finding buyers; recovering 50–60% after two years is a reasonable planning assumption, which cuts the effective cost per owned month by a third. Second, the "machine already exists" case: if the PC would be there anyway, the true comparison is marginal electricity versus rental — and then ownership wins at much lower utilization.
What the spreadsheet leaves out
- Failure risk is yours now. A used card has no warranty and a past. Budget a small probability that your $820 becomes a $820 lesson, and buy from sellers who test under load.
- Noise and heat are real costs. 350 W under load is a space heater with fans. Apartment dwellers in August already know this; everyone else discovers it the first summer.
- Your hours count too. Cloud is someone else's 2 a.m. hardware failure. Home is yours. If your time is worth anything, add a fudge factor.
- Latency, privacy and offline are worth something. Not every benefit shows up in the cost column. A local model that answers about your private documents has value no rental receipt captures.
What we'd actually do
If you're under ~3 hours a day, rent and stay liquid. Between 4 and 8 hours, buy the used 3090 if the machine exists and you'll keep it two years or more — the payback is real but not dramatic. Above 8 hours, stop reading comparison posts and buy the card. And whatever you choose, re-run the numbers when your workload changes: utilization is the whole game, and utilization is the thing most likely to move. The spreadsheet didn't get simpler in 2026 — it just finally fits on a fridge magnet.
Comments