Vast.ai

Vast.ai is an open compute marketplace that rents out third-party GPUs by the second: it pools 20,000+ GPUs from 1,000+ independent hosts into a live quoted-price pool, matches supply and demand through bidding, and delivers everything as Docker containers. It does not build foundation models of its own, and it does not own GPU assets — what it builds is quote ranking (DLPerf and reliability scoring), the host daemon and container orchestration, Secure Cloud datacenter certification, and Serverless scheduling. The name is easy to confuse: it is unrelated to enterprise storage vendor VAST Data and to Shanghai-based AI chip company Vastai Technologies.

How billing works Pure pay-as-you-go, no subscription. You top up credits before starting an instance, with a $5 minimum; usage is billed per second and never rounded up, and instances stop automatically once the balance runs out. Vast.ai says it does not mark up host prices, so the same GPU model is listed by many hosts individually and the listed price is the price you pay. Standard rentals require no KYC — signing up with an email address is enough to start an instance — and compliance and content responsibility rest with the tenant.

Choosing between the three instance types On-Demand is priced by the host, comes with guaranteed resources and is never preempted, so what you buy is certainty. Interruptible is a bid market: Vast.ai's own wording is more than 50% cheaper, a bid below min_bid is simply rejected, and when an on-demand request displaces you the instance is paused rather than destroyed, with data preserved and work resuming automatically, which suits fine-tuning and batch jobs that can restart from a checkpoint. Reserved is up to half off, but it can only be converted from an already running On-Demand instance and cannot be ordered directly. There is also Serverless: deploy a model as an endpoint, pay the same per-second rates with no serverless markup, and pay only for storage while an endpoint is Idle.

Hidden costs Storage is billed for as long as an instance exists: the meter keeps running in Stopped state, and only destroy truly stops it — the most common way newcomers lose money, with observed tiers from about $0.0015/hr to $0.0074/hr. Bandwidth is set by the host: the observed median is $2.67/TB, and the highest seen is $53.33/TB. The website shows a live starting price that moves hour by hour (captured on 2026-10-19), so the from price and the median for the same model can differ by a factor of two.

Who it suits, and who it does not It suits researchers, students, and independent developers doing LoRA fine-tuning, batch inference, diffusion model image and video generation, or benchmarks, and teams that need an H200 or B200 for a short window to validate large VRAM footprints. It does not suit customer-facing services that need a 99.9% SLA, heavily regulated core systems, or teams without ops skills that expect something that works out of the box: the shared conclusion of Vast.ai and third-party reviews is that it is good for running jobs and bad for running services, so checkpoints are not optional but mandatory. Accept the capability limits too: a single container does not support Docker-in-Docker, there is no built-in monitoring dashboard or CI/CD, and multi-node InfiniBand clusters are thin on the ground.

How it differs from RunPod and Lambda For the same model at on-demand rates, the Vast.ai median is still below RunPod Secure Cloud (H100 SXM at $2.40/hr on the Vast.ai site versus $3.99/hr on RunPod), the GPU catalog is broader and the cheap tail is more extreme; the cost of that is wide variance in supply quality, so filter reliability to 0.98 or above before comparing prices. RunPod feels more like a finished product, with smoother templates and Serverless; Lambda sells whole eight-GPU machines plus 1-Click Clusters out of its own datacenters — a training-cluster business, so the same $/GPU/hr buys a different configuration and the two are not directly comparable. If you need compliance isolation, go with Secure Cloud (certified third-party datacenters, HIPAA BAA available), at roughly 1.7 to 3.3 times the lowest market price.

Editorial take Vast.ai leads on three things — price, GPU model coverage, and API completeness — and is the compute marketplace most worth listing right now; the deductions are all for the lack of an SLA, reliability variance, and hidden billing. Come in expecting the cheapest AWS and you will be disappointed.