Analytics cookies
I use Google Analytics to see which pages are useful and which are not. If you accept, it stores cookies in your browser; if you decline, it only gets anonymous data, with no cookies. Cookie policy
Doing the sums
This is the page for whoever is deciding with money on the table. You give it your workload and it tells you, per month, what each road costs: buying the card and paying for electricity, renting a graphics card by the hour, or paying tokens to an API. Sometimes the cloud wins, and we say so.
Guide prices. Last checked in October 2026. They can vary from shop to shop and from month to month. In 2026 memory and graphics cards have become much more expensive (an RTX 5090 has gone from about €2,400 to over €6,000) and prices are not expected to fall soon.
Your case
From 1,000 to 5,000,000 a month (logarithmic scale).
What you pay per kilowatt-hour, taxes included.
Working in parallel gets more done, but it needs spare memory.
By default, the same model hosted in the cloud (there you do not pick the quantisation: the provider does).
We spread the purchase price across those months.
Per month
—
—
Break-even
It pays for itself from
—
—
At your current volume
—
—
Head to head
—
| Where | € / 1,000 | Against doing it yourself |
|---|
What does not show up on the invoice
Three of them argue for keeping the card at home. The fourth argues against. All four are true.
Payslips, contracts, medical records, invoices with names and account numbers: with the card in your own machine, the text never crosses the internet and never lands in anybody's logs. With an API, every document travels to a third party, and somebody has to decide whether that is allowed.
The card has no requests-per-minute limit and does not grind to a halt on a Tuesday afternoon because half the world is using the same model. If the work arrives in bursts —month-end, 30,000 documents at once— the only thing in charge is your own machine.
Once the card is bought, your cost per document depends only on electricity. API prices go up, go down, change name, and sometimes the model you were using is retired. A model you have downloaded is still there in three years, just as fast and at the same price.
A card at home makes noise, gives off heat and needs looking after: updating the engine, watching that it does not hang, replacing it if it dies. And above all: the frontier models in the cloud are more capable than a 27B running at home in 4 bits. If your job needs that level of reasoning, the euro comparison above is beside the point.
Where every euro on this page comes from.
API prices were looked up in September 2026 on each provider's own pages, standard rate, no batch or cache discounts, converted to euros. They change without warning: each one's note is in the table.
Card prices are from that same month, and that month they were through the roof: demand for AI memory has eaten the consumer market. Each card carries its note (new, second hand, list price). Buy at another moment and the write-off number changes, not the reasoning.
Speed figures always carry their label: for the combination you have picked right now. Renting by the hour assumes you only pay while the card is working; in practice you pay for the whole machine, so it is an optimistic floor, not a ceiling. The formulas are in the method.
The calculation engine (tw-engine.js) is not there yet. Reload the page in a moment.