Rent out your GPU to AI developers while you sleep
AI developers all over the world rent other people's graphics cards by the hour, and Vast.ai is the biggest open marketplace for it. Do the real maths first, with UK electricity and a calculator that tells you if your card will pay. Then build the host: Ubuntu Server, the NVIDIA driver, Vast's software, and the settings that decide whether you get rented.
- The habits that keep reliability above 90%: maintenance windows, alerts, a UPS and no surprise reboots
- A kit: profit calculator (Python and spreadsheet), preflight script, parts list and a weekly hosting log
- An NVIDIA GPU with more than 7GB of VRAM (Maxwell or newer). The ones that rent best are 24GB and up: RTX 3090, 4090, 5090
- Wired broadband with 500Mbps upload and download, and a public IPv4 address (not CGNAT)
- A spare SSD of 200GB or more for renters' containers, and a machine you won't use for anything else
- Comfortable installing Linux and using SSH. New to it? Start with Terminal survival kit
- Curious what renters do on the other side? See Rent a cloud GPU
Find out if your GPU will pay before you plug anything in
Run the numbers with your card, your tariff and real market prices, so you build the version that pays.
How the money worksOn Vast.ai you set a price per GPU hour and renters pay by the second while they use it. Vast says it plainly: the rate you set is the rate you receive, with no commission taken from your earnings; Vast adds its own margin on the renter's side (vast.ai/hosting ↗). Payouts go out weekly through Wise, PayPal or Stripe once you've built up $20 (payout docs ↗).
What sells is VRAMRenters pick by memory first, because that decides which models fit. A 24GB card (RTX 3090, 4090) or a 32GB card (RTX 5090) is the floor for most AI work. To see what renters actually pay for each card, check the GPU grid at vast.ai/pricing ↗ or the price table in Rent a cloud GPU. Your host price has to sit below what renters see.
The cost that decides it is electricityA rented RTX 5090 machine pulls around 700W at the wall, all day. At the Ofgem price cap average of 26.32p per kWh (Direct Debit, check the current cap ↗) that's about 18p an hour just in power. Download gpu_pays.py (standard library only) and run it:
The RTX 5090 example (new build, £2,900)python gpu_pays.pyOutputHost price $0.45/hr load 700W idle 80W power £0.2632/kWh hardware £2,900 over 36 months rented income kWh power after power after hardware 25% £62 172 £45 £17 -£64 50% £124 285 £75 £49 -£31 70% £174 375 £99 £75 -£5 90% £224 466 £123 £101 £20 Power cost of one busy hour: $0.244. Set your interruptible minimum bid near this. Break-even after hardware: rented 75% of the time at this price.Pick the winning setupA brand-new £2,900 rig on a domestic tariff needs to be rented about three quarters of the month, so treat that as the expansion move, not the starting point. The strongest start is a card you already own, say an RTX 4090 that sits idle most of the day. With no hardware to pay off, it's in profit from the first rented hour:
A 4090 you already ownpython gpu_pays.py --price 0.30 --load-w 550 --idle-w 70 --hardware 0Outputrented income kWh power after power after hardware 25% £41 139 £37 £5 £5 50% £83 226 £60 £23 £23 70% £116 296 £78 £38 £38 90% £149 366 £96 £53 £53 Power cost of one busy hour: $0.191. Set your interruptible minimum bid near this. Break-even after hardware: rented 19% of the time at this price.And the same new 5090 rig on cheaper power, for example a business tariff or a large solar array (
--unit-rate 0.15), breaks even at 50% rented instead of 75%. Power price is the biggest lever you have.Careful- What it takes. Hardware: an NVIDIA card with 24GB+ VRAM to compete, in a machine you'll dedicate to hosting (personal use fails verification).
- Internet: 500Mbps upload and download on a wired connection with a public IPv4 address.
- Power: your tariff decides the profit; run the calculator with your real unit rate.
- Hours: an afternoon to build and install, then a few minutes a week, plus fixing problems quickly when they happen.
- Time to first rental: depends on your card, price and verification; popular cards priced at or below the market go fastest.
- Who this suits: someone who already owns a high-VRAM card or has cheap power, likes running Linux, and wants hardware to pay its own way.
Everything's in the kitDownload the kit: the calculator (Python and spreadsheet), the preflight script, a parts list and a weekly log.
If it breaksHow do I know my real watts? + 1 more
How do I know my real watts?
--load-w and --idle-w.The £ sign prints as a strange character on Windows
$env:PYTHONIOENCODING="utf-8" in PowerShell first, or open the spreadsheet version instead.