Review basis
What we could verify
We checked Vast.ai's current pricing, billing, machine-tier, rental-type, and security documentation on 4 September 2026. The price table below is a dated marketplace snapshot, not a promise of today's rate. This is a source audit, not a hands-on performance benchmark, so we do not assign an invented score.
Affiliate link. Marketplace terms vary by host.
What is Vast.ai?
Vast.ai is a GPU marketplace rather than a fixed-price cloud. Hosts list machines, and renters select an individual offer by GPU, region, network, price, reliability, and other hardware details. The result is broad choice, but two listings for the same GPU can differ materially.
The marketplace includes consumer GPUs such as RTX 3090 and RTX 4090 as well as datacenter accelerators such as A100 and H100. Prices change with supply, demand, region, machine specifications, and host reliability.
Machine tier matters. Vast.ai labels new or untested machines as Unverified, machines that pass its internal checks as Verified, and qualifying certified datacenter offers as Secure Cloud. For production or sensitive work, that distinction is more useful than choosing the lowest hourly number.
Interruptible, On-demand, and Reserved
Interruptible instances use a direct bidding system. If another renter places a higher bid, or an on-demand rental takes priority, the interruptible instance can be paused. Running processes stop, so important work must be saved to disk and designed to resume.
On-demand instances have high priority and a fixed host-set price. They are the safer fit for time-sensitive jobs, although the offer's maximum rental duration still needs to cover the planned workload. Reserved instances are prepaid on-demand rentals with a longer commitment and a possible discount.
Checkpointed batch work can make good use of interruptible capacity. Production inference, a deadline-driven training run, or any job that cannot resume should start with on-demand or reserved offers instead.
Why the GPU Rate Is Not the Full Cost
Vast.ai's official pricing documentation separates compute, storage, and bandwidth. Compute is billed while the instance runs, storage continues while an instance exists even when stopped, and transfer prices vary by host. The offer card exposes the breakdown when you inspect its price details.
The July snapshot below shows why a single "from" rate is weak planning data. Availability and price can move quickly, and a low compute rate may come with different reliability, network, location, storage, or maximum-duration terms. Compare the complete offer and keep a portable copy of important data.
Vast.ai Marketplace Snapshot (29 July 2026)
| GPU | VRAM | Cheapest listing | Median listing | Best For |
|---|---|---|---|---|
| RTX 3090 | 24 GB | $0.05/h | $0.15/h | Experiments, SD |
| RTX 4090 | 24 GB | $0.13/h | $0.36/h | SD XL, fast inference |
| A100 SXM4 80GB | 80 GB | $0.13/h | $0.90/h | 70B fine-tunes |
| H100 PCIe | 80 GB | $1.33/h | $2.00/h | Fastest single-GPU inference |
| H100 SXM | 80 GB | $1.73/h | $3.49/h | Multi-GPU training |
Both price columns were read from vast.ai/pricing on 29 July 2026. They are a historical snapshot, not a current quote. Check the live offer card for compute, storage, bandwidth, reliability, machine tier, and maximum duration before renting.
Vast.ai Pros & Cons
- Individual offer details before rental
- Verified and Secure Cloud filters
- Interruptible option for resumable work
- Compute billed by actual usage time
- Storage charges continue while stopped
- Bandwidth prices vary by host
- Interruptible jobs can be paused
When to Use Vast.ai
Vast.ai is a practical fit for: price-sensitive, fault-tolerant batch jobs; experiments that save state frequently; and teams willing to compare individual hosts and total offer costs.
Use more caution for: production inference, regulated data, jobs with a hard deadline, or workloads that need identical hardware across repeated runs. Secure Cloud and Verified filters reduce risk, but they do not remove the need for backups and application-level security.
Vast.ai Alternatives
- RunPod Community Cloud: compare its current marketplace terms and templates when you want a different setup experience.
- Lambda Labs: compare current dedicated-cloud availability when consistency matters more than the lowest listed rate.
- Paperspace: consider it when a managed notebook workflow is more important than machine-level marketplace choice.
- CoreWeave: evaluate it for larger managed and multi-node deployments.
Primary Sources and Review Method
This review was checked against Vast.ai's official documentation rather than copied from marketplace summaries. The pricing snapshot remains dated because live offers change continuously.
- Pricing and cost components
- On-demand, reserved, and interruptible instance types
- Machine tiers and offer-card fields
- Container isolation and Secure Cloud guidance
- Minimum deposit and account-credit behavior
Verdict
Vast.ai is worth checking when price and machine choice are the main constraints and the workload can tolerate marketplace variability. The sensible decision is made offer by offer: prefer Verified or Secure Cloud for important workloads, compare the total cost rather than the GPU rate alone, confirm maximum duration, and keep checkpoints plus external backups.
Vast.ai FAQ
Is Vast.ai safe?
Vast.ai isolates clients in unprivileged Docker containers, but its own security documentation says provider security varies. For sensitive workloads, use the Secure Cloud filter, encrypt data before upload, keep credentials outside the instance, and delete the instance when the work is complete.
What is the minimum you need to spend to start on Vast.ai?
The official quickstart lists a $5 minimum deposit and supports credit card, BitPay, or Crypto.com. When the balance reaches zero, Vast.ai automatically stops running instances. Enable autobilling and a low-balance alert if the job cannot tolerate that stop.
What is the difference between interruptible and on-demand instances?
On-demand instances have high priority and a fixed host-set price. Interruptible instances use direct bidding and can be paused when another user outbids you or an on-demand rental takes priority. Use interruptible capacity only for jobs that save state and can resume; use on-demand for time-sensitive work.
How do I find the best host on Vast.ai?
Start with Verified machines or Secure Cloud for production use. Then compare the reliability score, DLPerf, network speeds, maximum rental duration, storage price, and bandwidth price shown on each offer card. The cheapest compute rate is not necessarily the lowest total cost.
Does Vast.ai have H100?
Vast.ai lists H100 offers when marketplace hosts have capacity available. Search the live dashboard for the required H100 variant, GPU count, region, reliability, network, and maximum rental duration before planning a job.
How does Vast.ai compare to RunPod Community Cloud?
Vast.ai exposes individual machine offers with host, reliability, network, storage, and bandwidth details. Compare those complete offer terms with RunPod at the time you rent. A lower GPU rate can be offset by data transfer, storage, setup time, or a shorter maximum rental duration.