Compute Economics
Local GPU vs. Cloud Break-even
Is a workstation you own cheaper than renting cloud GPUs, at the hours you will actually run it?
A local box is a purchase; cloud GPUs are a bill. Both are priced per GPU-hour, at the hours you will really run and the years you will really keep it.
What this calculator answers, what you need, and what to do with the result
Have this in front of you
- The all-in hardware price: card, host, tax and shipping.
- Monthly hours of real GPU use, taken from your logs rather than the busy week.
- The cloud rate you would otherwise pay, per GPU-hour, at your commitment level.
Words on this page, in plain English
- GPU-hour
- One GPU running for one hour. The unit both the cloud bill and the owned machine are priced in.
- Utilisation
- The share of the month the hardware is actually computing. Idle hours are free on cloud and billed locally.
- Residual value
- What the card is worth used when you sell it. Counting it as zero makes owning look worse than it is.
What to do with the answer
- 01Count a realistic 24-month life. A card that is obsolete in 18 months changes the answer more than the purchase price does.
- 02Add the hours you spend maintaining drivers and cooling. They are not free, they are just unpaid.
- 03If utilisation is under about a third of the month, rent. Above that, the hardware tends to win.
Compute capital · local vs. cloud
Local Workstation vs. Cloud GPU
Cost per GPU-hour at the utilisation you will actually run — purchase price, power, maintenance and residual value against the on-demand rate, over the life you will actually keep the box.
01The hardware
02How you will use it
Owned cost includes purchase minus residual value, electricity at the load wattage you enter, and maintenance; it excludes your own setup and debugging time, which is real but personal. Cloud cost excludes egress, storage and the premium for reserved vs spot capacity. Hardware is assumed replaced at end of life at the residual value stated.
Method
How this number is calculated
- 01
Owned cost, net of resale
Hardware price minus what it is worth at the end of its life, plus power, plus maintenance. At $7,500 with 15% residual value the net capital outlay is $6,375.
- 02
Power over the whole life
Kilowatts at the wall multiplied by total hours and your electricity rate. Total hours are monthly hours times the life in months, so power scales with use while depreciation does not.
- 03
Cost per GPU-hour, both sides
Owned cost divided by total hours, compared with the cloud rate multiplied by the same hours. Comparing a capital purchase to an hourly rate without normalising hours is the step that makes local hardware look cheap.
- 04
Break-even hours per month
Fixed owned cost divided by (life in months times the cloud rate minus power cost per hour). That is the utilisation above which ownership is cheaper, and the number to hold your estimate of real usage against.
Not priced: your admin time, hardware support, and the inability to scale past one box during a deadline. Electricity assumes a residential or small-office rate; colocation adds rack and bandwidth costs.
Worked numbers
Pre-computed scenarios
One $7,500 workstation, 3-year life, 15% residual value, 450 W at the wall, $0.18/kWh, $200/yr maintenance, against $1.10/hr cloud. Only the hours per month change.
| Scenario | Inputs | Outcome |
|---|---|---|
| Light use, one box | 120 h/mo · $7,500 box · $1.10/hr cloud | $1.70/hr owned vs $1.10/hr cloud — rent cloud gpus by $2,573 over 3 years |
| Steady use, one box | 320 h/mo · $7,500 box · $1.10/hr cloud | $0.69/hr owned vs $1.10/hr cloud — buy local hardware by $4,764 over 3 years |
| Heavy use, one box | 560 h/mo · $7,500 box · $1.10/hr cloud | $0.43/hr owned vs $1.10/hr cloud — buy local hardware by $13,568 over 3 years |
Questions
What people ask before trusting this number
At what monthly utilisation does a local GPU pay for itself?
Break-even is the fixed cost of the card divided by the hours you run, net of the electricity you stop buying per cloud hour. At $7,500 with 15% residual value, $0.18/kWh power and a $1.10/hr cloud rate, that is roughly 190.1 hours a month. Below it the cloud bill buys the same compute for less; above it ownership wins.
Should power draw be measured at the wall or on the card?
At the wall, for the whole box. A 450 W card sits in a system that pulls 550 to 700 W under load once the processor, memory and fans are counted, and the meter bills the wall figure. Using the card rating alone understates owned cost by roughly a third.
Why does residual value change the answer so much?
Because it is subtracted before anything else is divided. A card that resells for 40% of its price costs 60% of the headline figure to own, which pulls break-even hours down by a third against a card that resells for nothing. Datacentre cards hold value differently from consumer cards, so use the price you could actually sell at.
The box sits idle most of the month. How is that charged?
Two ways. Power is charged only for the hours you run, because idle draw is small. The hardware is charged across its whole life regardless, because idle silicon still depreciates. That asymmetry is why the model reports break-even hours per month rather than one total.
What is missing from this comparison?
Your own admin time for driver updates, failed runs and hardware support, and the cost of being unable to scale past one box during a deadline. Both push the decision toward renting. They are left out because they vary too much to price, so add them by hand before treating a narrow win as a win.
Embed this calculator on your site ↓
Bloggers & authors: Embed this interactive calculator directly in your articles or comparison guides. It loads dynamically via a responsive iframe (auto-resizing to prevent double scrollbars) and provides a direct, high-value tool for your audience.
<div class="giniloh-calculator-embed" data-calculator="gpu-breakeven"></div>
<script src="https://giniloh.com/embed.js" async></script>