Vast.ai

What it is for

Vast.ai is a GPU marketplace: hosts post offers, you rent on-demand, interruptible, or reserved, and you pay the live bid by the second. Use it when the job is pick a priced offer, not when you need a single vendor SKU.

It is not open source as a product. There is no monthly seat. An H100 SXM offer on the public feed started at $1.734/hr when we copied it.

What makes it unique: the cover is Vast.ai’s official Create / offer chrome — a live GPU row (type, reliability, $/hr) you actually rent.

Key features

• Create / rent — filter GPU, CUDA, disk, and reliability on the live offer list, then Rent and SSH or open Jupyter (cover; Create; Find and rent).
• On-demand, interruptible, reserved — guaranteed box, preemptible bid, or a 1 / 3 / 6 month term on the same marketplace (Instance types; Pricing).
• Templates — pick a prebuilt Docker image or save your own so the next GPU boots the same stack (Templates; Choosing a template).
• Volumes — attach a local disk on the same host (/data); it survives a destroyed instance so you can re-rent that machine and remount it (Volumes).
• Serverless endpoints — autoscale inference workers (vLLM / TGI / Comfy) without holding a VM (official Serverless intro; Serverless).

Who it's for

Best fit: engineers who will read the offer row (reliability, bandwidth, $/hr) before they click Rent.

Poor fit: teams that want one published SKU from a single operator (Lambda).

Pricing

Verified 2026-10-22 (Asia/Shanghai) from the official public feed (updated 2026-10-22T05:30:16Z, gpu-pricing-public.json) plus vast.ai/pricing. Bids move every hour. Copy the offer you will rent.

Live feed examples (USD / GPU / hr)

Interruptible is cheaper and can be reclaimed. Bandwidth is often extra on the console row. No published monthly seat.

Limitations / costs

• The cheap min is not the box you get if that offer is gone.
• Interruptible can vanish mid-job. Checkpoint.
• Host quality varies. Read reliability / verification on the row.
• Not a Lambda SKU. Support and networking are marketplace-grade.

Alternatives

• Runpod — published Secure Cloud rows; H100 PCIe $2.89/hr.
• Lambda — 8× H100 SXM $3.99/GPU/hr if you want one operator.
• Modal — functions if you do not want to pick a host.

Official tutorial

Official Vast AI channel video: Autoscaling GPUs for AI Inference: Introducing Vast.ai Serverless (oEmbed author: Vast AI).

Written path: Pricing plus Create.

Website

https://vast.ai

Source

Catalog row curated for AI Tools Nav. Cover is Vast.ai’s official Create / offer chrome. Live feed 2026-10-22T05:30:16Z: H100 SXM from $1.734/hr. Tutorial: Serverless on @vast_ai (oEmbed).

Last verified

2026-10-22 (Asia/Shanghai) editorial check.