Image and video generation
Run ComfyUI, Stable Diffusion, or other diffusion workflows without buying a card for occasional projects.
01 / Start with the workload
Pick the machine around the work.
The right GPU depends on what you are running, how long it needs to stay available, and how much data you need to move. Start with the task, then compare offers.
LLM inference
Try open models, test serving configurations, and compare memory requirements across GPU classes.
Fine-tuning
Choose enough GPU memory for your model and context, and consider how long the training run should remain uninterrupted.
Transcription and audio
Process a batch of recordings with GPU acceleration, then release the compute when the queue is finished.
Rendering
Move a project render to remote GPUs when your local workstation is busy or the deadline is tight.
Development and testing
Bring up a temporary environment for CUDA, libraries, and prototypes before committing to infrastructure.
02 / Know the tradeoffs
Choose how much certainty you need.
Vast offers different rental types. The lowest displayed hourly price is not automatically the best fit for a workload that must finish on time.
| Type | Useful when | Tradeoff to understand |
|---|---|---|
| On-demand | You want a straightforward rental with higher priority. | Compare the full offer price and host details before launching. |
| Interruptible | Your task can tolerate being paused or interrupted. | Lower cost may come with less continuity; avoid it for time-critical work. |
| Reserved | You can plan ahead or commit funds for a longer run. | Check prepayment and reservation terms against your actual usage. |
03 / Before you launch
Look past the GPU name.
Marketplace offers can differ in ways that matter to both performance and the final bill. Check these before you commit.
Memory and configuration
Confirm GPU memory, GPU count, CPU, system memory, disk, and the software template you plan to use. A familiar model name alone does not guarantee your workload will fit.
Reliability and access
Review host reliability, location, network, connection method, and rental type. For important runs, decide whether interruption is acceptable before optimizing for price.
Total cost, not just hourly GPU cost
Prices change with supply, demand, host, and configuration. Storage may continue billing while an instance is stopped, and bandwidth can add cost. Check current rates and remove resources you no longer need.
04 / Referral disclosure
Use my link to explore Vast.ai.
Referral disclosure: I may receive referral credits from Vast.ai if you create an account using my link and purchase credits. The link does not provide a discount or change your price.
Open Vast.ai offers
This link opens Vast.ai's GPU marketplace in a new tab.
05 / Common questions
Good to know before signing up.
Does the referral link lower my price?
No discount is promised by this referral link. It is a way to attribute an eligible new client account to the referrer under Vast's referral program.
Do I need to keep an instance running?
No, but understand which costs continue. Vast's documentation says storage can remain billable while an instance is stopped. Review the current billing details and remove resources when you are finished.
Are GPU prices fixed?
No. Vast is a marketplace where hosts set offer prices and rates vary with supply, demand, GPU configuration, rental type, and other offer details. Use the live pricing page when comparing.