11 November 2026 · VRAMfarm team
Rent or buy an RTX 4090 or RTX 5090 for AI: break-even hours in 2026
A used RTX 4090 cost about €2,000 in Europe in November 2026, and a new RTX 5090 usually €4,000–5,000, roughly double its list price of about €2,100. Renting the same cards costs from €0.59 and €0.76 an hour. With those numbers, buying a 4090 starts paying off at around 90 hours of GPU work a month if you already have a suitable PC, and around 175 hours if you have to build one. For a 5090 at street price the line sits at about 190 and 275 hours. Every input is listed, so you can redo the sum with your own prices.
What a card really costs to own
The purchase price is only the first line. The figures below are rounded estimates for someone in the EU in November 2026.
| RTX 4090 | RTX 5090 | |
|---|---|---|
| VRAM | 24 GB GDDR6X | 32 GB GDDR7 |
| Memory bandwidth | 1,008 GB/s | 1,792 GB/s |
| Board power | 450 W | 575 W |
| Card price, Nov 2026 | about €2,000, used (new stock has largely gone) | about €4,000–5,000 new; the ~€2,100 list price is rarely in stock |
| PC around it, if you need one | about €2,000 | about €2,000, with a 1,000–1,200 W power supply |
| Whole system under load | about 0.6 kW | about 0.75 kW |
| Electricity per hour at €0.30/kWh | about €0.18 | about €0.23 |
| Resale after 2 years (assumed) | about half | about half |
A few notes behind the table:
- Prices moved a lot in 2026. Memory shortages pushed GDDR7 cards far above list price, and 2025–26 also saw used 4090s resell close to what they cost new. Check a price comparison site on the day you decide.
- The PC is not free. A 4090 wants an 850 W or bigger power supply; a 5090 wants 1,000 W or more, a case with room for a long card, and ideally 64 GB of RAM so models can load without swapping. RAM was expensive in 2026 too.
- Electricity depends on where you live. €0.30 per kWh is a round European figure; German households often pay more, around €0.40.
- Resale is a guess. Half after two years is a middle assumption. A new generation or the end of the memory shortage can push used prices down quickly.
The break-even calculation
The method is short: work out what owning costs per month, then compare it with renting the same hours.
Owning, per month = (purchase price − resale value) ÷ months of use + electricity per hour × hours.
For a 4090 card in a PC you already own: (€2,000 − €1,000) ÷ 24 ≈ €42 a month, plus €0.18 per hour of work. If you also buy the PC, add another (€2,000 − €1,000) ÷ 24 ≈ €42, so €84 a month.
For a 5090 at street price: (€4,500 − €2,250) ÷ 24 ≈ €94 a month, plus €0.23 per hour, or about €136 with a new PC.
Renting, per month = hours × the hourly price. No fixed part.
RTX 4090: monthly cost by hours of use
| GPU hours per month | Rent (€0.59/h) | Own, card only | Own, card + new PC |
|---|---|---|---|
| 20 | €11.80 | about €46 | about €88 |
| 50 | €29.50 | about €51 | about €93 |
| 100 | €59.00 | about €60 | about €102 |
| 200 | €118.00 | about €78 | about €120 |
| 400 | €236.00 | about €114 | about €156 |
RTX 5090: monthly cost by hours of use
| GPU hours per month | Rent (€0.76/h) | Own, card only | Own, card + new PC |
|---|---|---|---|
| 20 | €15.20 | about €99 | about €141 |
| 50 | €38.00 | about €106 | about €148 |
| 100 | €76.00 | about €117 | about €159 |
| 200 | €152.00 | about €140 | about €182 |
| 400 | €304.00 | about €186 | about €228 |
The rental column shows today’s price, so it stays current; the ownership columns use the November 2026 assumptions above. The point where the lines cross is the fixed monthly cost divided by (rental price per hour − electricity per hour). With the catalogue prices of November 2026 that gave:
- RTX 4090: about 90 hours a month (card only) or 175 hours (with a new PC).
- RTX 5090 at street price: about 190 hours or 275 hours.
- RTX 5090 at list price (~€2,100), if you can find one: about 90 hours, card only.
Put into a day: 90 hours a month is about 3 hours of real GPU work every day; 175 is about 6; 275 is about 9, every day, for two years.
What the simple sum leaves out
On the renting side, add setup minutes after the server is ready (installing packages, downloading models; the 2–5 minutes of launch before that are free), the 15-minute minimum per launch, traffic for downloading models, and storage if you keep a saved folder: 50 GB of models costs 50 × €0.03 a month. For most people that is a few euros. On the owning side, add the time you spend on drivers and CUDA versions, the heat (a 600–750 W heater in your room in July) and noise, and the risk of a used card without warranty. There have been reports of used 4090s sold online with the GPU chip removed, so buy used from someone you can check.
When buying is clearly better
- Daily heavy use. If the GPU is busy most days for several hours (long training runs, an image or video generation business, a model you serve all day), the fixed cost spreads thin and owning wins, often by a lot at 300+ hours a month.
- You already have the PC. The card-only column is the one that matters, and it halves the break-even point.
- You also use it for games or 3D. Part of the price is then not an AI cost at all.
- You found a card near list price. A 5090 at about €2,100 moves its break-even down to the same 90 hours as a used 4090.
- Your data may not leave your premises, by contract or by choice, or you need to work offline.
If two or three of these are true for you, buy the card.
When renting wins
- Bursty use. A weekend project, one LoRA a month, a client job twice a quarter. At 20–50 hours a month, the rental bill is a fraction of the ownership cost, and you pay nothing in the months you do nothing.
- You sometimes need 48–80 GB. A 32 GB 5090 is still short of what some jobs need: Wan 2.2 14B video only just fits and is comfortable on 48 GB, a 70B language model in 4-bit needs about 43 GB of weights plus room for context, and QLoRA on a 70B model wants 48 GB or more. A 48 GB workstation card costs well above a 5090, and an 80 GB data-centre card many times more. Renting one costs from €0.72 an hour for an RTX A6000 with 48 GB or €1.46 for an A100 with 80 GB. If you are not sure what your model needs, the VRAM picker shows it.
- No upfront cash. You start with a top-up from €10, billed per minute, and unspent paid balance can be refunded to the original payment method on request.
- You want to test before you buy. Rent the card you are considering for a month and look at how many hours you actually used. The number is easy to overestimate in advance.
- You want to switch generations. When a faster card arrives, you pick it in the catalogue instead of selling the old one.
The hybrid that many people end up with
Owning and renting are not either-or. A common setup:
- A card at home for daily work. The GPU you already have, or a 16–24 GB card. Prompt tests, small LoRAs, debugging a training script, all without a meter running.
- Rented memory for the big jobs. When a run needs 48–80 GB or would take your home card all night, launch a bigger GPU for those hours and stop it when the job ends.
- The same environment on both sides. Keep your models and scripts in
/workspaceand use the saved folder between rentals, or build your own Docker image that you can run at home and launch on a rented server. The VRAMfarm API can start a server for an overnight batch and stop it when the batch is done.
To find your real hours, check your billing history after a month of renting, or log a few weeks of home use with nvidia-smi --query-gpu=utilization.gpu --format=csv -l 60 >> gpu.log. Count only the hours the GPU was actually busy; those are the hours that go into the tables above.
If your number comes out under about 90 hours a month, rent. If it is well above 175 for a 4090 and you can live with the heat, buy. In between, the card you already own plus rented memory for the heavy days is usually the cheapest of the three. Prices for every card are on the pricing page, and the RTX 4090 and RTX 5090 pages list current machines.