What is the difference between on-demand and interruptible on Vast.ai?

On-demand instances are guaranteed to keep running until you stop them, at a fixed hourly rate. Interruptible instances run on a bidding system and can be reclaimed when someone outbids you, so they cost far less but are only safe for work you can checkpoint and resume. Use interruptible for dataset generation and batch upscaling, on-demand for serving an API.

Why does reliability score affect price?

Vast.ai is a marketplace of independent GPU hosts. Each host has a reliability score based on uptime and performance history. More reliable hosts can charge a premium because their instances are less likely to drop. Filtering for a higher minimum reliability raises your effective hourly rate but reduces failed jobs.

Is Vast.ai cheaper than RunPod or Lambda Labs?

Often yes, especially on interruptible instances and consumer GPUs like the RTX 4090, because it is a competitive open marketplace. The trade-off is variability — uptime, bandwidth, and disk speed differ per host. For predictable production workloads RunPod Secure Cloud or Lambda Labs reserved capacity may be safer.

Does this include disk and bandwidth charges?

No. Vast.ai bills disk storage and internet bandwidth separately from GPU compute, and rates vary per host. This estimator covers compute hours only — add a buffer for storing large model checkpoints and transferring outputs.

Vast.ai GPU Cost Estimator

Vast.ai GPU cost estimator

Vast.ai is an open marketplace where independent hosts rent out GPUs, so prices swing with supply and demand and with each host’s reliability score. The two levers that decide your bill are the pricing mode — on-demand (guaranteed) or interruptible (bid-based and cheaper) — and how many hours per month you run. This estimator turns those choices into a monthly number.

How it works

Select a GPU and the tool loads representative on-demand and interruptible rates. A reliability multiplier nudges the rate up when you demand more dependable hosts, because those hosts charge a premium. Monthly cost is simply the effective hourly rate times your expected hours. The estimator shows both pricing modes so the savings from accepting interruptions is explicit.

Tips for the marketplace

Use interruptible for batch work. If your pipeline checkpoints and resumes, interruptible instances can cut compute spend by half or more.
Set a sensible reliability floor. Going too low saves pennies but costs you in failed runs and re-launches; a mid-to-high floor is usually the sweet spot.
Watch disk and bandwidth. Large checkpoints (Flux, SDXL fine-tunes) and heavy output transfer are billed separately and can rival compute on small jobs.
Compare before committing. For steady production load, price a reserved Lambda Labs node against Vast.ai on-demand — predictability sometimes wins.