Guides / Computing costs and local AI
Is renting out a spare GPU or CPU worth it?
How power draw, electricity price, utilization, payout and platform fees combine into net income per hour and a break-even date for a rented GPU or server.
Practical guide
Several online marketplaces let you rent out a gaming GPU or a spare server to strangers who need compute by the hour. The pitch is simple: the hardware is sitting there anyway, so why not let it earn something? Whether it actually earns anything depends on a handful of numbers you can estimate before you sign up: how much power it draws, what your electricity costs, how often it is rented, what renters pay and what the platform keeps. This guide shows how those numbers combine, using the same method as our GPU and CPU hosting calculators. Nothing here is a promise of income; market rates change constantly and the examples use made-up but plausible inputs.
The method
The calculation splits every hour into two cases: the machine is either rented, running at full load, or idle, waiting for a renter while still drawing some power.
- Power cost per rented hour = load power in watts ÷ 1,000 × electricity price per kWh × PUE.
- Power cost per idle hour = idle power in watts ÷ 1,000 × electricity price × PUE.
- Average power cost per hour = u × rented cost + (1 − u) × idle cost, where u is utilization, the share of hours the machine is rented.
- Average revenue per hour = u × payout per rented hour × (1 − platform fee).
- Net per hour = average revenue − average power cost.
- Net per month = net per hour × 730 hours (8,760 hours in a year ÷ 12).
- Break-even = hardware cost ÷ net per month, if net per month is positive.
PUE stands for power usage effectiveness, a metric defined by The Green Grid. It multiplies the power the computer draws to cover cooling and power conversion. A machine in a spare room with no extra cooling is about 1.0 to 1.2. If you run an air conditioner harder to get rid of the heat, use a higher number. The industry-average data center PUE reported by the Uptime Institute's annual survey is about 1.5 to 1.6.
For a GPU host, load power is the board power of each GPU times the number of GPUs, plus the rest of the system (CPU, memory, fans and power supply losses), which the calculator treats as constant whether the machine is rented or idle.
Worked example: two RTX 4090 cards
The inputs below are illustrative. Board power is NVIDIA's specification for the card; idle power is a rough estimate, and the payout is an illustrative figure, not a current market rate.
| Input | Value |
|---|---|
| GPUs | 2 × GeForce RTX 4090 |
| Board power per GPU | 450 W |
| Idle power per GPU | 25 W |
| Rest of the system | 120 W |
| Utilization | 50 % |
| Electricity price | $0.16 per kWh |
| PUE | 1.1 |
| Payout | $0.35 per GPU-hour, before fees |
| Platform fee | 25 % |
| Hardware cost | $3,600 |
- Load power: 2 × 450 + 120 = 1,020 W. Idle power: 2 × 25 + 120 = 170 W.
- Power cost per rented hour: 1.02 kW × $0.16 × 1.1 ≈ $0.1795.
- Power cost per idle hour: 0.17 kW × $0.16 × 1.1 ≈ $0.0299.
- Average power cost: 0.5 × $0.1795 + 0.5 × $0.0299 ≈ $0.1047 per hour.
- Payout per rented hour after the fee: 2 × $0.35 × 0.75 = $0.525. Averaged over all hours: 0.5 × $0.525 ≈ $0.2625.
- Net per hour: $0.2625 − $0.1047 ≈ $0.158.
- Net per month: $0.158 × 730 ≈ $115. Net per year: about $1,382.
- Break-even: $3,600 ÷ $115.18 ≈ 31 months.
The machine uses about 478 kWh a month at this utilization, which is worth comparing against your current bill.
What moves the answer
Change one input at a time and the break-even point moves a lot:
| Change | Net per month | Break-even |
|---|---|---|
| As above | $115.18 | 31.3 months |
| Utilization 25 % instead of 50 % | $46.67 | 77.1 months |
| Electricity $0.32 instead of $0.16 | $38.73 | 92.9 months |
| Payout $0.20 instead of $0.35 | $33.05 | 108.9 months |
Halving the utilization cuts net income by more than half, because the idle power is still being paid for. Doubling the electricity price cuts it by about two thirds. A payout drop of 43 percent cuts it by more than 70 percent, because the power cost does not drop with it. Each of these pushes break-even beyond six years.
The calculator also reports a break-even utilization: the share of rented hours at which revenue exactly covers power. With the original inputs it is about 8 percent. That sounds reassuring, but it only covers electricity, not the hardware.
When it loses money
Take one RTX 3090 (350 W board power, about 30 W idle, 80 W for the rest of the system) at a low payout of $0.10 per hour with a 25 percent fee, rented 30 percent of the time, on electricity at $0.35 per kWh with a PUE of 1.1.
- Power cost per rented hour: 0.43 kW × $0.35 × 1.1 ≈ $0.166. Payout after the fee: $0.075. Every rented hour loses about $0.09 before counting idle time.
- Average net per hour: about −$0.057, or about −$41 a month.
No utilization fixes that, because each rented hour costs more in power than it earns. The only fixes are cheaper power, a higher payout or not hosting at all.
Worked example: a 64-core server
CPU marketplaces often pay per core-hour. Take an AMD EPYC 7763 server: 64 cores, 280 W rated processor power, 120 W for the rest of the server under load and 110 W at idle. With utilization at 60 percent, electricity at $0.15 per kWh, a PUE of 1.1, an illustrative payout of $0.01 per core-hour and a 20 percent fee:
- Power cost per rented hour: 0.40 kW × $0.15 × 1.1 = $0.066. Per idle hour: 0.11 kW × $0.15 × 1.1 ≈ $0.018.
- Payout per rented hour: 64 × $0.01 × 0.80 = $0.512.
- Net per hour: 0.6 × $0.512 − (0.6 × $0.066 + 0.4 × $0.018) ≈ $0.307 − $0.047 = $0.260.
- Net per month: about $190. On $3,500 of hardware, break-even is about 18 months.
Halve the payout to $0.005 per core-hour and net falls to about $78 a month, with break-even at about 45 months.
What the calculation leaves out
- Hardware wear and failure. Thousands of hours at full load can shorten a GPU's life, and fans and power supplies wear out. The calculation assumes the hardware lasts.
- Resale value. Break-even treats the hardware cost as money to recover. If you would have owned the card anyway, or can sell it later, the real picture is better than that; if it loses value quickly, it is worse.
- Internet and other costs. Some platforms need a fast, stable connection, a static address or a minimum uptime. A separate internet plan is a monthly cost.
- Taxes. Rental income is usually taxable. The calculation is before tax, and none of this is tax advice.
- Your time. Setup, updates and support requests take time that does not appear in the numbers.
Before you buy hardware to host
Measure the machine's real power draw with a plug-in wall meter, both idle and under load. Find your all-in electricity price: total charges on the bill divided by the kWh used, including delivery. Look at current listings on the platform you would use, because payout rates move weekly. Then run the numbers at a pessimistic utilization. If the result only works at the optimistic end, it is a gamble rather than a plan.
Sources
- U.S. Energy Information Administration. Electric Power Monthly, Table 5.6.A — Average Price of Electricity to Ultimate Customers by End-Use Sector. www.eia.gov/electricity/monthly/
- Uptime Institute. Global Data Center Survey (annual). Industry-average power usage effectiveness (PUE). uptimeinstitute.com/
- The Green Grid (2012). WP#49 — PUE: A Comprehensive Examination of the Metric.
- NVIDIA, AMD and Intel product specification pages (total graphics power / board power, TDP / default TDP, core counts).