GPU Prices Went Up Again — Where Does Buying Stop Making Sense and Renting Start?
I went to look at prices, thinking it was about time I replaced my graphics card, and stopped. They have gone up. Not by a little, either.
Running local AI on your own machine takes a graphics card. And that starting assumption looks like it is coming apart. If so, does renting one by the hour become the sensible option?
How much difference is there between buying and renting? How much speed do you lose when you rent? Which costs are easy to miss? Here is where it stood on 6 August 2026.
This article is mostly research. Where a number comes from my own hardware, I say so. I have not actually rented anything yet. The plan is to rent and measure after this article, and that plan is at the end.
The previous article in this series:
https://kimono-oyaji.com/en/rent-a-gpu-cost-math-en/What is happening to graphics card prices?
The clearest case is the RTX 5090. It launched at a suggested price of $1,999. As of August 2026 it is $4,288. More than double.
| When | What happened |
|---|---|
| January 2026 | A 10–15% increase |
| May 2026 | A further increase, on the RTX 5090 alone |
| Most recently | 20–30% across all models (the third increase in 2026) |
It is not only NVIDIA. AMD and Intel have raised prices, and so have the games console makers.
AI data centres are buying memory in volume, and there is not enough left over for consumer products. The memory makers have raised their prices too, and the shortage is expected to run until around 2028.
So this does not look like something that comes down if you wait.
From the point of view of running local AI, this is not somebody else’s problem. Large models need a card with a lot of memory, and memory is exactly what has gone up the most.
What does renting cost?
Here are the hourly services, per hour.
| What you rent | RunPod | Vast.ai |
|---|---|---|
| A100 (40GB) | $0.60 | $0.52 |
| A100 (80GB) | $0.79 | $0.67 |
| L40 (48GB) | $0.69 | $0.31 |
| RTX 4090 (24GB) | $0.69 | — |
| H100 (80GB) | $2.89 | — |
Vast.ai comes out cheaper. Anyone can list their own hardware there, which brings the price down. There is a catch, and I come back to it below.
The question I had was this: buying has doubled, so what has renting done? As far as I can tell, rental prices have climbed much more gently. But that is the picture right now. Data centres are competing for the same memory, so it seems more natural to assume it reaches the rental side eventually.
How much slower is a rented GPU?
A GPU you rent is usually sharing a machine with other people. The worry is what that costs you. Looking into it, the answer flips completely depending on how it is rented out.
Reports of 15–25% slower
One measurement puts it at 65–76% of what the card can do
Measured difference under 1%
(4,250 against 4,210 tokens per second)
Both services I looked at rent out as containers. On that basis, the speed itself should not drop much.
But that is about speed once a generation has started. From what I have measured here, most of the waiting was never in that part.
On my own machine, reading a 202GB model off disk takes 31 minutes (110MB per second). On a rented machine that read happens over the network. The waiting that happens before anything runs should matter more when renting, not less. This is the part I want to measure for real.
The costs that are not on the rate card
Look only at the hourly figure and the sums will not add up later. These are the ones that struck me as dangerous not to know.
- Stopping does not stop the storage bill. Charges continue until you delete it. Leaving 200GB sitting there for ten days works out at roughly $7–10 on its own
- Data transfer is billed too. About $2.50 per 100GB. Simply pulling down the weights of a large model adds up
- There are reports of being billed while the host machine was down, with no refund given
- Which is why some people put the real bill 20–40% above the listed price (for hosts who have not been vetted)
Pick the cheaper option and you take on more of this uncertainty. Cheap and dependable trade against each other fairly directly.
Buy or rent: not “how many hours" but “for what"
Last time I did this calculation I assumed eight hours a day and wrote that the two lines cross at about 2,000 hours. Let me correct that. I do not use it eight hours a day. In practice it is a few dozen minutes, whenever something occurs to me.
So I redid it by hours per month: how many years before buying a used RTX 3090 costs the same as renting the equivalent continuously. I put the purchase at ¥160,000, about $1,070 (the going rate for a used card as of August 2026; these are one-off items, so shop and timing move it).
| Hours per month | Roughly what that is | Years to break even |
|---|---|---|
| 10 hours | A short session two or three times a week | 30–71 years |
| 30 hours | About an hour a day | 10–24 years |
| 100 hours | Three hours-plus a day | 3–7 years |
| 200 hours | Most days, for long stretches | 1–4 years |
Putting numbers on it broke an assumption of mine. At hobby levels of use, it never pays for itself. Thirty or seventy years is not a statement about the lifespan of the card; it means you will not use it that much.
· It is there the moment you think of something (renting puts sign-up and loading ahead of you)
· Your data never leaves the house
· Forget to switch it off and all you lose is electricity
· You learn by fiddling with it (getting a setting wrong does not cost you by the hour)
Which also means: if you do not need those four, renting is the reasonable choice.
Split by what you are doing, it came out like this
| What you are doing | Better suited | Why |
|---|---|---|
| Daily lookups, summarising, coding companion | At home | Short sessions, many of them. Renting means more waiting than working |
| Trying a model too large for your machine | Rent | It will not run at home at all. A few times a month costs very little |
| Batch-generating images or video | Either | Running for a few hours straight costs a couple of dollars to rent |
| Handling data that cannot leave | At home | Not a question of money |
| Something that runs all the time | At home | Renting means never stopping equals never stopping the bill |
When would you rent for days at a time?
So far this has all been by the hour, but renting for days on end is also a thing. For an individual, it came down to three cases.
- Teaching a model your own habits (fine-tuning). Your own card does not have the capacity, and it takes time. On a large rented card it can be over in a few hours to a day
- Processing a lot of things at once. Translating hundreds of articles, converting a pile of video. One at a time is fine at home; multiply it and you are running overnight
- Running dozens of variations to compare. Repeating the same measurement while changing settings. The benchmarking I do falls into this
This is where the hourly rate starts to matter. Three days straight is 72 hours.
| What you rent | Per hour | 72 hours (3 days) |
|---|---|---|
| RTX 3090 (cheap host) | about $0.13 | about $9 |
| RTX 3090 (pricier host) | about $0.30 | about $22 |
| RTX 4090 class | about $0.69 | about $50 |
Nine dollars for three days is not a bad price for one attempt at fine-tuning. Provided it finishes as calculated, that is. Fail partway and re-run it and the total climbs. And forget to stop it and you are billed while you sleep.
Multi-day renting is also the hardest kind to predict a bill for. Set your ceiling in advance, is the lesson.
Next I will actually rent one and measure
Everything above is research. Whether the numbers hold is something you only find out by renting. Here is what I plan to measure.
| What to measure | Why |
|---|---|
| Generation speed on my machine and on a rented one, same model, same settings | Whether “barely any slower" is true |
| Total time from renting to the first answer coming back | How much the loading wait costs you when renting |
| The itemised bill afterwards | How much storage and transfer added |
| The same conditions twice, at different times of day | How much other people’s usage moves it |
Matching the conditions to my existing measurements is the essential part. Without that, a comparison does not mean anything.
Mentioned in this article
The card I used as the buying example. Prices have risen, but it is still a realistic way to have 24GB on your desk.
NVIDIA GeForce RTX 3090 24GB24GB VRAM, runs 27B-32B
As an Amazon Associate we earn from qualifying purchases.
The hourly rental services. I have not actually rented from them yet (this article is research only). These are referral links, so a referral benefit comes to me. It does not raise the price for you.
In summary
- Graphics card prices rose three times in 2026, and the RTX 5090 has more than doubled. The cause is the memory shortage, which is expected to run until around 2028
- Rental prices have climbed more gently so far. But they are competing for the same memory, so it is more natural to assume it arrives eventually
- On speed, there is a measurement putting container-based rental within 1%. Both services I looked at work that way
- But the loading before anything runs should weigh more heavily when renting (202GB took 31 minutes here)
- Stopping does not stop the storage bill. Transfer is billed too. Some put the real total 20–40% above the listed price
- Working out payback in hours does not hold at today’s prices. At 30 hours a month it takes 10–24 years
- The reasons to own one are not financial: instantly available, data stays home, forgetting to stop only costs electricity, and you learn by fiddling
- Renting for days on end suits fine-tuning, bulk processing, and running dozens of comparisons. Three days is $9–50, but forget to stop it and you pay while you sleep
What renting is really for turned out to be less about money than about touching a size that will not fit on your own machine. Next time I rent one and find out how close these sums are.
Sources
Prices in this article were checked on 6 August 2026. Where a yen figure appears it is Japanese retail, converted at ¥150 to the dollar. Both rates and exchange rates move.