Cloud GPU rental costs less than owning a card when the card sits idle. A ¥36,000 4090 rig, spread over 3 years, is about ¥1.64/hr ($0.23) at an impossible 100% load, including ¥0.32/hr of power. At 30–50% real use, the worked hour costs about $0.45–$0.75, close to the $0.49 RTX 4090 rental median.
01. Which is cheaper, cloud GPU rental or a card on the desk?
Rental is cheaper per finished hour whenever the card would be idle.
Cloud GPU rental bills the hour the card works. A card on the desk was paid on arrival. The socket keeps drawing on empty nights. The renter pays the busy hour. The owner pays the quiet ones too.
A finished frame on a rented 4090 costs that hour's listing. The same frame on a bought rig also carries the hours between jobs. When quiet hours are the normal month, rental is the lower cost per frame you ship.
Owning wins only on hours you actually run, for years.
A full-time 4090 hour in this model is about $0.23. It appears when the rig is powered and working close to every hour. The purchase in this model is spread over three years. A desk that renders most days, for years, can land under the rental median. A card bought so the team has one does not.
Own the hours you will run. The $0.23 hour is the full-clock case. Quiet hours do not inherit it.
The median, not the floor, is the rental number in this comparison.
Vast.ai lists an RTX 4090 median of $0.49/hr. The $0.14 floor is one listing. The GPU rental cost per hour to budget is the median. The floor can be gone when the job is ready. It is the wrong number for a month of work.
The two listings sit about 3.5× apart. That multiple belongs to the 4090 on Vast.ai. Teams that rent GPU cloud hours, and teams that rent cloud GPU hours, should type $0.49 for a 4090.
Side by side, the commitment is the point.
| Way you pay | What you commit | Cost at 100% power-on | Cost per hour you actually run, at 30–50% use |
|---|---|---|---|
| Buy a ¥36,000 4090 rig | Cash up front + power | about $0.23/hr (about ¥1.64) | about $0.45–$0.75 |
| Cloud GPU rental, RTX 4090 | The rented hour only | $0.49/hr median | $0.49/hr median (idle risk sits with the host) |
| Cloud GPU rental, the floor listing | The rented hour, if you can get that machine | $0.14/hr | $0.14/hr, and you should not budget this |
At 100% power-on, the bought rig is about $0.23/hr. The rented median stays $0.49/hr. At 30–50% use, the bought worked hour is about $0.45–$0.75. The rented median stays $0.49/hr, because the idle risk sits with the host. Leave the floor row out of the budget.
02. What is inside the $0.23 owned hour?
About ¥1.28 of the hour is the card and rig, spread over three years.
The cash in this model is a whole rig at about ¥36,000. That is the card plus the platform that has to exist before the card can run. Spread across three years of powered-on time, the hardware slice is about ¥1.28 an hour. ¥1.28 already is that spread. Use it as given.
About ¥0.32 of the same hour is electricity on a 600W rig. The 600W figure is the card and the rest of the machine at the wall. That ¥0.32 is the China residential case. The regional table is next.
Together that is about ¥1.64, or $0.23.
Keep the trio intact. About ¥1.28 of hardware, about ¥0.32 of power, about ¥1.64 an hour. That ¥1.64 is $0.23. Those three are one statement. When you need the full-time owned hour, say $0.23.
100% power-on is a ceiling, not a studio calendar.
Nights, renders that fail, and days between projects are the normal case. A 100% clock means the machine earns in every hour of the three years, including sleep and rejected shots. Studios do not live on that clock. Treat $0.23 as the ceiling of the model. The worked hour, after idle time, is the number that decides cloud GPU vs buying a GPU.
03. Why does a rented 4090 median sit near $0.49?
Power alone is $0.10 to $0.22 per hour for a US or European host.
A 600W rig in China is $0.047/hr. The same box does not cost the same to leave on. Residential rates diverge, and the hour of electricity diverges with them. A host in the United States pays $0.102 for that hour of power. A host in Germany pays $0.222. The EU average and the UK sit between them, at $0.152 and $0.198. The span from $0.10 to $0.22 is that spread, read off the rows, before any recovery of the card.
The host still has to recover the card.
Electricity explains the gap. It does not set the listing by itself. A host who bought a 4090 still has to get the card back from the hours someone rents. Power changes by country. The card still has to come back over the life of the machine. A China power cost of $0.047 does not become a $0.047 rental ad. A US power cost of $0.102 is not the whole $0.49 median. The listing has to cover the card after the kilowatt-hours are paid.
A floor price does not include a host who can stay at that price.
Budget $0.49. The $0.14 floor can be a spare machine, or a short listing from a host whose books already differ from yours. You cannot see those books from the sticker. You can see the middle of the board. Vast.ai's 4090 median is $0.49/hr. That is the month you can plan.
Power for a rig at about 600W, card plus platform:
| Region | Residential rate | Whole rig per hour | Versus China |
|---|---|---|---|
| China | ¥0.532/kWh ($0.079) | $0.047 (¥0.32) | 1× |
| United States | $0.17/kWh | $0.102 | 2.2× |
| EU average | $0.254/kWh | $0.152 | 3.2× |
| United Kingdom | $0.33/kWh | $0.198 | 4.2× |
| Germany | $0.37/kWh | $0.222 | 4.7× |
China is the base row: $0.047 an hour, or ¥0.32, marked 1×. The United States runs 2.2× that, Germany 4.7×. These rows are power. They are not rental prices, and they do not turn the $0.14 floor into a budget.
04. How does utilization flip the answer?
At 50% use, the owned worked hour is about $0.46.
Half the hours on, half off. The hardware spread still assumes the full clock you used at purchase. The hours that ship the work carry the quiet ones. At 50% use, that worked hour is about $0.46. It already meets the rental median of $0.49. Owning and renting are the same bill at that load, before a dead fan or a failed render pushes the owned hour up.
At 30% use, the owned worked hour is about $0.77.
A weeknight calendar often looks like this, or quieter. The worked hour is about $0.77. Renting at $0.49 is the lower bill. The card on the desk is still wearing through the hours you did not render. Cloud GPU rental stops when you stop.
The band to remember is $0.45–$0.75.
It is an illustration, not a survey of studios. The 30% and 50% rates are planning figures for this model. About $0.46 at half use and about $0.77 at 30% use are the readings behind the band. Remember $0.45–$0.75. Use $0.23 only for a rig that essentially never rests. Use $0.49 when you would rather hand the idle hours to a host on Vast.ai.
05. What else does ownership add that the hour-rate hides?
A dead fan, a driver week, and a machine you cannot take on a trip.
Those hours are paid and not billed to a client. A dead fan takes the card down with it. A driver week is a week the card exists and the shot does not ship. A tower stays home when you travel. The client never sees those hours. You already paid for them.
Rental adds queue risk and setup time.
A shell you configure yourself can burn the first hour before a single frame. You pay the live rate while a weight download finishes and a path gets fixed. The machine behind a floor screenshot can be gone when you return. The hour you actually get is often the median. If you rent a GPU in the cloud, idle-card risk sits with the host. Setup risk on a raw shell stays with you.
Task rental removes the setup hour for image and video.
You trade shell access for a finished file. For image and short video, that file is the whole job. Task rental skips the shell. The job goes in. The files come back.
SpareHive is not in the live price table. The planning band sits roughly 50–60% under the $0.49 RTX 4090 median, list price unset, rentals not open. Join the waitlist.
06. Which choice fits which month?
Rent if this quarter's work is a burst.
A launch, a film week, a catalog. The card would sit idle on either side. Pay the $0.49 median for the hours inside the burst, then stop. A ¥36,000 rig through the quiet months does not make that burst cheaper.
Buy if the card will earn its keep most days for three years.
Otherwise the $0.23 figure never arrives. Buying fits a desk that renders most days, across those three years, on work the card can hold. If you cannot name those days, you are living in the $0.45–$0.75 band and pricing a ceiling you will not hit. In that band, the bought hour does not clear the rental median by enough to ignore fans, drivers, and setup.
Pick the card before you pick the billing mode.
The hour is only a saving if the card matches the job. If you are still matching memory to the work, start with which NVIDIA GPU for AI. If you already know you will rent, read NVIDIA GPU rental prices and budget the median of the card you will get. Then own or rent a 4090 from the first table, using the 30–50% column, not the floor.
FAQ
Is cloud GPU rental always cheaper than buying?
No. A 4090 rig run almost full-time lands near $0.23/hr in this model, under the $0.49 rental median. At 30–50% use, the worked hour is about $0.45–$0.75, and rental wins. Full-time power-on is the ceiling of the model, not the calendar to plan on.
Why not budget the $0.14 RTX 4090 listing?
No. That floor is about 3.5× under the $0.49 median. It is one machine, not the price a team can plan a month on. Vast.ai lists both numbers. The median is the one that repeats.
Does cheap electricity in China change the buyer's math?
No. It changes the host's cost. A 600W rig is $0.047/hr to power in the China row of the table, and $0.102–$0.222 in the US and Europe. The renter still pays the listing.
Is a $5 top-up the same as owning a card?
No. A small top-up is a way to test a rented hour. SpareHive's planned $5 top-up is not live, and it does not replace a card you intend to run for years.