Vast.ai pricing: what you really pay, and a fixed-price alternative paid in crypto
How Vast.ai's marketplace prices work, what is billed besides the GPU, two worked bills, and a fixed-price alternative paid only in crypto. Petaflare is a fixed-price alternative and our own offer, so we say plainly where Vast.ai costs less.
Vast.ai pricing is set by each host, not by Vast.ai. You pay the host's GPU price per second while the instance runs, plus disk storage for as long as the instance exists, even when it is stopped, plus bandwidth for every byte sent or received. On 18 November 2026, the median price in Vast.ai's public price feed was $1.33 an hour for an H100 SXM, $0.76 for an A100 SXM4 and $0.36 for an RTX 4090. Interruptible rentals are often half the on-demand price or less, and reserved ones up to 50% off.
Vast.ai GPU prices on 18 November 2026
Median price per GPU-hour from Vast.ai's public price feed, the number of offers the feed listed as available, and our fixed price for the same card.
| GPU | Vast.ai median | Offers available | Petaflare | Petaflare vs median |
|---|---|---|---|---|
| RTX 4090 | $0.36 | 687 | $0.34 | 5% less |
| RTX 5090 | $0.46 | 922 | $0.49 | 7% more |
| L40S | $0.53 | 91 | $0.87 | 66% more |
| A100 SXM4 · PCIe | $0.76 · $0.51 | 128 · 67 | $0.99 | 31% more than SXM4 |
| H100 SXM · PCIe | $1.33 · $2.00 | 36 · 2 | $2.39 | 80% more than SXM |
| H200 · H200 NVL | $3.50 · $3.17 | 35 · 113 | $3.29 | 6% less than H200 |
| B200 | $7.00 | 73 | $4.89 | 30% less |
Source: Vast.ai public GPU price feed, updated 19 November 2026 at 09:30 UTC, cited with attribution. A median is the middle offer: half the hosts ask less. The H100 PCIe median rests on two offers. Our prices are on price per GPU-hour.
Three ways to rent on Vast.ai
Vast.ai's pricing documentation describes three kinds of rental.
| Type | How it works | Price |
|---|---|---|
| On-demand | Fixed price set by the host, resources guaranteed while you pay | The host's listed price |
| Interruptible | Lowest priority: the instance may be paused for higher-priority renters | Often 50% or more below on-demand, according to Vast.ai |
| Reserved | Pay in advance for a period | Up to 50% off, according to Vast.ai |
Source: Vast.ai pricing documentation, read on 19 November 2026.
What a Vast.ai instance is billed for
The GPU price is only one of three lines on the bill, and each host sets all three.
| Item | When it is billed | Price set by |
|---|---|---|
| GPU and machine | Per second while the instance runs | The host |
| Disk storage | Continuously while the instance exists, running or stopped, until you delete it | The host |
| Bandwidth | Per byte transferred, upload and download | The host |
Credits must be added before an instance starts, and running out of credits stops your instances. Source: Vast.ai billing reference.
Two example bills
Disk and bandwidth prices differ from one offer to the next, so the bills below use assumed host rates: $0.10 per GB a month for disk and $10 per TB for bandwidth. Check the rates on the offer you pick. Both jobs use one H100 SXM at the $1.33 median and a 200 GB disk.
One H100: a busy week and an occasional month
| Busy week | Occasional month | |
|---|---|---|
| GPU time | 40 hours: $53.20 | 10 hours: $13.30 |
| Disk, 200 GB | Kept 7 days: $4.67 | Kept 30 days: $20.00 |
| Bandwidth | 170 GB: $1.70 | 50 GB: $0.50 |
| Vast.ai total | $59.57 | $33.80 |
| Per GPU-hour used | $1.49 | $3.38 |
| Same GPU time on Petaflare | $95.60 | $23.90 |
Petaflare sells hours in blocks: the busy week needs four 12-hour H100 blocks ($114.72, 8 hours left on the account), the occasional month one block ($28.68, 2 hours left). Unused hours never expire. Assumes the job fits on our machine's local disk.
Reading the two bills
For a busy week, Vast.ai is clearly cheaper: the H100 median is so low that disk and bandwidth barely move the total. For occasional use, the disk kept between sessions becomes the biggest line, and the real price per GPU-hour more than doubles. That is where a fixed price with no storage line wins. The break-even depends on the card: the gap is widest on the H100, A100 and L40S, where the medians are far below our prices, and reversed on the B200 and RTX 4090.
Vast.ai prices move from week to week
The median for the same card, on three dates a fortnight apart.
| GPU | 20 October | 4 November | 18 November |
|---|---|---|---|
| H100 SXM | $1.80 | $1.50 | $1.33 |
| H200 | $3.65 | $4.32 | $3.50 |
| A100 SXM4 | $0.44 | $0.55 | $0.76 |
| L40S | $0.35 | $0.35 | $0.53 |
| RTX 4090 | $0.40 | $0.35 | $0.36 |
| RTX 5090 | $0.58 | $0.51 | $0.46 |
Source: Vast.ai public GPU price feed, daily medians, cited with attribution. The L40S median rose 50% in two weeks, while the H100 SXM fell 26% in a month.
Vast.ai for Stable Diffusion
Image generation is where Vast.ai's consumer cards shine. On 18 November the feed listed 358 RTX 3090 offers at a $0.16 median, 687 RTX 4090 offers at $0.36 and 922 RTX 5090 offers at $0.46. A 24 GB RTX 3090 runs SDXL and Flux with FP8 weights; we do not rent that card at all. Our RTX 4090 at $0.34 sits just under Vast.ai's median, and our RTX 5090 at $0.49 just over it.
Which card you need depends on the model and the resolution: see Stable Diffusion VRAM needs before you pick a host.
When Vast.ai is the better choice
- Busy, short projects on data-center cards. Its medians for the H100, A100, L40S and H200 NVL are below our fixed prices.
- Older or consumer cards. The RTX 3090 and many other models we do not rent.
- Restartable batch work. Interruptible rentals cut the price again.
- Long commitments. Reserved rentals take up to 50% off.
- Paying only for what you run. Per-second billing, with no block of hours to buy.
For a side-by-side with RunPod, the other popular marketplace, see the RunPod and Vast.ai comparison.
A fixed-price Vast.ai alternative paid in crypto
Petaflare rents seven NVIDIA cards at one fixed price each, with no host-by-host variation and no separate disk or bandwidth line on our price list. A stopped machine uses no GPU-hours and keeps its disk. We are cheaper than Vast.ai's median on the B200, the H200 and the RTX 4090, and dearer on the others.
Payment is crypto only, sent to us directly, with no card on file and no identity check. Vast.ai takes crypto through BitPay and Crypto.com and does not refund credits bought that way; our hours never expire. The trade-offs: our smallest block costs $25.50 to $29.70, and a person prepares each machine, so it does not start in seconds. If you are also weighing RunPod, see what changes instead of RunPod, and for every provider, the provider comparison.
Vast.ai pricing questions
Is Vast.ai a good company?
We will not grade a competitor. The facts that matter: it is a marketplace, so you rent from independent hosts whose machines and reliability vary; credits must be bought before use; disk is billed while an instance is stopped; and credits bought with crypto are not refunded.
How much is 1 GPU hour?
It depends on the card. On 18 November 2026, Vast.ai's medians ran from $0.16 for an RTX 3090 to $7.00 for a B200. Our fixed prices run from $0.34 for an RTX 4090 to $4.89 for a B200, and an H100 costs $6.88 to $12.29 at the big public clouds.
How much does it cost to rent an RTX 5090 per hour?
The Vast.ai median was $0.46 on 18 November 2026. Ours is $0.49, or $0.44 on the 3,000-hour block; RunPod charged $1.39 on Secure Cloud and $0.69 on Community Cloud.
Does Vast.ai bill storage when an instance is stopped?
Yes. Its billing reference says stopping an instance does not avoid storage costs; only deleting the instance stops them.
Sources
- Vast.ai public GPU price feedread 19 November 2026
- Vast.ai pricing documentationread 19 November 2026
- Vast.ai billing referenceread 19 November 2026
- RunPod pricingread 19 November 2026