Guides / Computing costs and local AI

Utilization: the number that decides whether a rented GPU pays

What the share of rented hours does to one GPU's net income and payback, the three utilization thresholds that matter, and the payout trade-off.

Practical guide

By Steven Michael Barber · Published October 17, 2026 · Updated November 3, 2026 · 5 min read

Our hosting guide works through the whole calculation for a rented GPU: power in, payout out, fee, break-even. Every input in it is something you can look up or measure except two, which the market sets: the payout rate and utilization. Utilization, the share of hours the card is actually rented, also depends on how you compete, and it is the input that decides whether the venture pays. This guide holds everything else fixed, the payout included until the last section, and shows what utilization alone does to the answer.

What utilization is, and why it is the uncertain input

Utilization is rented hours divided by hours available. In the calculator's average 730-hour month, 438 rented hours means 60 percent utilization. Hardware power and your electricity tariff can be measured; demand, payout and how many hours a listing wins need market evidence. We have not measured utilization across hosting platforms. The table below shows the result at each level, not how likely that level is.

The calculator's example is one 350 W card (its RTX 3090 preset, from the vendor's specification) that idles at 30 W in a system that draws 80 W on its own, electricity at $0.15 per kWh (near the residential average the EIA publishes) with a 10 percent overhead for cooling (a PUE of 1.1 in the Uptime Institute's terms), a payout of $0.25 per GPU-hour, a 20 percent platform fee, and $900 of hardware. These are the calculator's undated example inputs, not current market rates; everything below uses them, and your own numbers belong in the calculator before any decision.

The whole range, in one table

The net is a straight line in utilization, rising by $10.75 a month for every ten points, because each rented hour earns the same $0.20 after the fee and costs the same $0.071 in power. The line does not start at zero: at 0 percent the host is paying $13.25 a month to keep an idle machine on. Payback is hardware cost divided by net, so its curve is steepest near the point where net becomes positive. A small utilization change near that point can change payback by years.

The three utilizations that matter

The calculator reports a break-even utilization, the point where revenue just covers power. Here it is 12.3 percent. That threshold excludes recovering the hardware cost and other expenses. Three illustrative budgeting targets show what different goals require:

• Paying the hardware back in two years needs $37.50 a month, which this card reaches at 47 percent utilization. In one year it needs $75 a month and 82 percent. The calculation does not tell you whether your listing can achieve either.
• Covering the card's own wear. A card run hard for rental does not last forever. Setting aside its price over a three-year life (an assumption; use your own) is $25 a month, and the host reaches that at 36 percent. Below that the host is slowly consuming the card for less than it costs to replace.
• Both at once, payback in two years and the replacement reserve, needs $62.50 a month, which is 71 percent utilization. That is a separate cash-allocation goal: recover the initial purchase and also build a fund for its eventual replacement. It is not an additional operating expense in the calculator.

With these inputs, about seven rented hours in ten meets the combined payback-and-reserve goal. A different goal, payout or power bill gives a different target. The table lets you compare those choices without assuming that any utilization is assured.

Payout and utilization pull on the same rope

The payout rate is the other market input, and it moves the whole table. At $0.20 per GPU-hour instead of $0.25, the two-year payback needs 65 percent utilization rather than 47, and the 60 percent host's net falls from $51 to $34 a month. At $0.15 there is no utilization at all that pays the card back within two years. A host who cuts the listed price to win more hours should check which side of that trade they are on: lowering the price from $0.25 to $0.20 is worth it only if it raises utilization by at least 18 points, from 47 to 65, just to stand still on payback, and by more than that to gain.

Worked example: checking the 60 percent row by hand

At 60 percent the card earns 0.6 × $0.25 × (1 − 0.20) = $0.120 an hour after the fee. Power costs (350 + 80) ÷ 1,000 × $0.15 × 1.1 = $0.071 for a rented hour and (30 + 80) ÷ 1,000 × $0.15 × 1.1 = $0.018 for an idle one, so the average is 0.6 × 0.071 + 0.4 × 0.018 = $0.050 an hour. Net is $0.120 − $0.050 = $0.070 an hour, which over 730 hours is $51.22 a month, and $900 ÷ $51.22 is 17.6 months to pay the card back. That is the row in the table, and the same three lines of arithmetic produce every other row.

Measure your own utilization before you believe any of this

After a full month, divide rented hours by the hours your card was available. Use that observed utilization with your real payout, fee and electricity price, and compare several months before treating it as a forecast. If the result misses your target, test what a price change or reduced idle time would do. Choose a review date and a maximum loss before starting, so the decision to continue follows your own budget rather than the illustrative thresholds here.

What this leaves out

The table counts electricity and hardware only. Heat and noise in the room, the wear reserve beyond the simple straight-line figure above, the risk that a platform changes its fee or its payout, the hours lost to updates and reboots, and the tax treatment of the income are all real and all outside it. The hosting guide's closing sections cover what to measure and what to check before buying hardware to host; nothing here is a forecast of what any platform will pay next month.