Buy or rent?
Is your own AI machine cheaper than paying per use? Fill in how your team would use it and see the monthly cost of each option.
Machine prices are estimates of European prices, converted to dollars without EU VAT, before US sales tax. Pay-per-use and rental prices are US dollar prices.Machine prices are estimates in euros, with VAT. Pay-per-use and rental prices are in dollars; we convert them at a recent exchange rate and add 21% VAT, like the machine prices.
Danylo Chernov (NoctHost) benchmarked five Ollama models on Hetzner Cloud CPUs and rents A16 and A40 GPUs by the hour. Self-reported, community measurement, not ours. Read Danylo Chernov's interview.
Payback against a $200 a month Pro plan
| Setup | Stored price | Monthly electricity | Payback | Month 6 | Month 12 | Month 24 | Against |
|---|---|---|---|---|---|---|---|
| Used RTX 3090 card-only | €1,246 / $1,395 | $5.94/mo US / €10.56/mo EU at a 250 W cap | 7.2 months | $1,431 vs $1,200 | $1,466 vs $2,400 | $1,538 vs $4,800 | vs $200/month Pro |
| Hybrid 3090 + $20/month Pro | €1,246 / $1,395 | $25.94/mo ($20 plan + $5.94 power) | about 8.0 months | $1,551 vs $1,200 | $1,706 vs $2,400 | $2,018 vs $4,800 | still paying $20/month |
| Full 3090 PC | €2,346 / $2,413 | $7.56/mo US / €13.44/mo EU | 12.5 months | $2,458 vs $1,200 | $2,504 vs $2,400 | $2,594 vs $4,800 | vs $200/month Pro |
The 7.2 and 12.5 month figures are a little longer than the dollar price divided by $200, so they leave room for electricity. The hybrid figure is about $1,395 divided by the $180 a month you stop paying while a $20 plan stays. Prices are the stored used-card median and our build total, checked 24 Oct 2026. EU electricity on the card is €10.56 a month at €0.32/kWh (33.00 kWh/mo, 250 W via nvidia-smi -pl 250), about 7.4 months in euros. A rented 24 GB GPU at $0.55/hr is $66/mo for 4 hours a day, and the card costs less after 23.2 months. Our estimate, not a bill.
Monthly cost for your team
20 people × 30 questions a day, in office hours, with Llama 3.3 70B. Change it in the form.
Cheapest for you: paying per use for Llama 3.3 70B, about $17about €18 a month.
To keep your data on your own machine, buying a used Mac Studio M4 Max 128GB costs about $112about €100 a month over 3 years, and pays back after about 31 months compared with renting a graphics card (Vast.ai L40S).
The Mac Studio M4 Max 128GB keeps up with your team: in a busy hour it answers about 6 people at once. Each answer then comes at about 4.5 tokens per second (tok/s), a bit slower than you read.
Buying new? The cheapest new machine we list that keeps up is the Mac Studio M5 Max 128GB, for about $168about €179 a month.
Where the numbers come from
- Exchange rate: 1 euro = 1.1195 dollars. This is the European Central Bank reference rate of 14 Oct 2026. Machine prices are European prices in euros, incl. VAT, excl. shipping; in dollars we take out the 21% EU VAT, before US sales tax. A used price has no VAT in it, so in dollars it is only converted. Pay-per-use and rental prices are US dollar prices before tax; in euros we add the same 21% VAT. Electricity is the price per kWh in the form; in dollars it is only converted.
- Machine: the Mac Studio M4 Max 128GB at a used-market price of 16 Oct 2026, an estimate not checked against a shop listing (unverified). Power: Apple's maximum for this Mac.
- Pay per use, same model: Llama 3.3 70B Instruct, list price of 15 Oct 2026: $0.59 per million question tokens and $0.79 per million answer tokens.
- Rented graphics card: Vast.ai L40S, $0.73 an hour, a typical marketplace price of 21 Oct 2026, for 176 hours a month. One card; a large team may need more than one.
- Top model: Claude Sonnet 5.5, list price of 14 Oct 2026: $2 per million question tokens and $10 per million answer tokens.
- Electricity: €0.3499 per kWh by default, Belgium, households, second half of 2025, taxes, levies and VAT included. Eurostat, read 20 Oct 2026. We count power only for the hours you chose, not idle power outside them. A Mac's figure is Apple's maximum for the whole machine; a PC's is its graphics card alone, so they are not like for like.
- Question sizes, in tokens (a token is about ¾ of a word): short 300 in and 200 out, medium 1,500 in and 500 out, long 6,000 in and 800 out.
- Keeping up: we compare the words your team needs per working hour with our estimate of how fast the machine writes, the same estimate as our hardware calculator. A machine can write several answers at once while their conversations fit in its memory, never more than there are people. We count the busiest hour as 2 times the average. A machine whose answers come slower than about 3.0 tokens per second (tok/s) is too slow for chat, even when it gets through the work. Reading long questions also takes time, which we do not count.
- A rented card's speed is not recorded, so we check it only roughly: it holds as many conversations as fit in its memory, and we count each answer as fast as on the fastest machine we list for this model with as many conversations at once. When even that is not enough for your team, we do not call one rented card the cheapest. We do not check the speed of a pay-per-use service.
- Not included: resale value, tax, financing and the time someone spends looking after the machine.
Costs and speeds are our estimate from list prices and hardware specs, not a quote or a measured test. Prices are dated and may change. How we calculate