Calculator

How long until a GPU pays for itself?

Buying hardware is capital plus electricity; an API is per token forever. This works out the month the two lines cross, from the list price and power draw we hold for the card and the live per-token price we sync for the model.

How this is worked out

Your numbers

Everything updates as you change it, and the address bar keeps up — so the answer you are looking at is the answer you can send someone.

50M tokens

Output tokens only — the ones you pay the most for. A busy support bot is around 50M.

40%

100% means flat out around the clock. Most self-hosted setups sit between 20 and 50.

Your bill, not our data — we do not track tariffs.

Self-host vs API break-even

1month

$1,499 of card plus $12.26 a month of electricity costs less than $1,250 a month of API calls from that point on.

Seeded, unreviewedCard list price
$1,499written 29d ago
Seeded, unreviewedRated power draw
350 Wwritten 29d ago
Synced from sourceAPI output price
$25.00per MtokCurrent22h ago
Output tokens
50 Mper month
Card busy
40%of the month
Electricity
$0.12per kWh — your assumption
2 of 3 inputs were written by us and never reviewed. The figure is exactly as good as those.
Ahead by, if you buy
GeForce RTX 3090 · Claude Opus 5

You are $1,499 down on day one and $5,927 ahead after 6 months. The line crosses zero where buying overtakes calling.

Line height = how much you are ahead by having bought the card. Below the zero rule the API is still cheaper.

-$1,499$357.60$2,214$4,071$5,9270mo2mo3mo5mo6moMonthsAhead by, if you buyBreak-even · 1 monthAhead byAhead by-$1,499Months: 0mo2 of 3 inputs were written by us andnever reviewed. The figure is exactlyas good as those.Ahead by-$261.26Months: 1mo2 of 3 inputs were written by us andnever reviewed. The figure is exactlyas good as those.Ahead by$976.47Months: 2mo2 of 3 inputs were written by us andnever reviewed. The figure is exactlyas good as those.Ahead by$2,214Months: 3mo2 of 3 inputs were written by us andnever reviewed. The figure is exactlyas good as those.Ahead by$3,452Months: 4mo2 of 3 inputs were written by us andnever reviewed. The figure is exactlyas good as those.Ahead by$4,690Months: 5mo2 of 3 inputs were written by us andnever reviewed. The figure is exactlyas good as those.Ahead by$5,927Months: 6mo2 of 3 inputs were written by us andnever reviewed. The figure is exactlyas good as those.
The hardware dataset
API, monthly
$1,250
Electricity, monthly
$12.26
Five-year saving
$72,765

Background

When buying a card actually beats paying per token

The decision looks like a price comparison and is really a volume question. An API charges you nothing until you use it and never stops charging; a card charges you everything on day one and then only draws power. Those two shapes cross at a volume, not at a price, which is why quoting one against the other per million tokens tells you almost nothing.

Two costs get left out of most comparisons and both favour the API. A card sitting idle still cost you its purchase price, so the break-even assumes you actually keep it busy — drop utilisation and the crossing moves out fast. And electricity is not a rounding error at 450 watts: a card flat out on a Danish tariff costs more per month to run than a modest API bill.

What this tool does not model: the machine around the card, your time, cooling, or the fact that a hosted model gets better while your card does not. Those all push the same way. Treat the break-even month as the earliest possible date rather than a forecast.

What this tool assumes

Compares
Purchase + power vs per-token API
Horizon
Five years
Card data
List price and rated draw, written by us
API data
Live output price, synced from OpenRouter
Not modelled
Host machine, cooling, your time

The data underneath

Local inference hardware

Every number this calculator used is a row you can open, with the source that reported it and the moment it was read. Nothing here is an estimate we could not show you.

Open the dataset →

Questions

What people ask

Why does it say the card never pays for itself?

At the volume you set, the API bill is smaller than the electricity bill plus the amortised purchase over five years. Raise the monthly tokens and the crossing appears — that is the honest answer, and it is the common one at low volume.

Is the list price what I would actually pay?

No. It is the launch list price we hold for the card, which is a reasonable anchor and rarely the street price. The trace list beside the result shows the figure used and when it was written, so you can substitute your own quote.

Why only output tokens?

Output is the expensive half and the half self-hosting displaces most clearly. Counting input as well moves the crossing earlier, so leaving it out keeps the answer conservative.

Where do the numbers come from?

Card price and power draw were entered by hand when the hardware dataset was built and have not been reviewed since; the model price is synced from OpenRouter. Every input beside the result says which it is and how old it is.