Vast.ai is a marketplace rather than a cloud. Rather than owning datacentres, it connects people who need GPUs with people who have them sitting idle, from hobbyists with a spare 4090 up to proper datacentre operators with spare capacity. The price consequence is dramatic. You can rent consumer and datacentre GPUs for a small fraction of what the hyperscalers charge, and for anyone learning, experimenting or running intermittent inference, that difference decides what is possible.
A student who cannot justify a hundred dollars for a weekend of experiments on a major cloud can absolutely justify eight. The documentation matches the product. It is practical and reasonably well organised, with an instances guide, templates documentation, a five minute walkthrough that does actually get you onto a GPU in about five minutes, CLI and Python SDK tutorials for automating provisioning, a serverless quick start, and a set of example workloads covering PyTorch, vLLM, ComfyUI and agent deployments that you can copy rather than assemble. Support is a 24/7 chat in the console plus a Discord.
There is no academy, no course catalogue and no certification, and to Vast's credit they do not pretend otherwise, which puts them ahead of several vendors whose learning centre turns out to be a help desk with ambition. What the docs do not cover well is the thing that actually distinguishes this platform, which is what it means to build on rented consumer hardware from unvetted hosts. The marketplace model has real consequences. Hosts have reliability scores and those scores exist because reliability varies.
An instance can go away because somebody's machine rebooted or their internet dropped or they decided to reclaim the card. Storage is tied to the instance in ways that will hurt you if you assumed otherwise. Throughput on a nominally identical GPU can differ substantially depending on the host's CPU, RAM and network. None of this is hidden exactly, and the documentation explains the mechanics, and there is not much guidance on the operational discipline it implies.
That guidance would be straightforward to write. Checkpoint aggressively and to external storage. Treat every instance as something that can vanish. Prefer high-reliability hosts for anything long-running and accept the premium.
Use the interruptible pricing for work that genuinely tolerates interruption and not for anything else. Benchmark before committing to a multi-day run. This is the actual curriculum of using a GPU marketplace well and it is left for users to discover, generally the hard way, twenty hours into an uncheckpointed training run. The privacy and trust question deserves a mention too, since you are running workloads on hardware owned by strangers.
For open model inference or a Kaggle experiment this hardly matters. For anything involving sensitive data it matters a lot, and the docs are quieter on that than I would like. As an educational resource this is thin, which is why the rating sits where it does. As a piece of infrastructure that materially widens who can afford to do this work, it is one of the more useful things in the ecosystem, and the growth in search interest over the past year suggests plenty of people have worked that out.
Three point seven for the documentation. The platform itself earns more than that if you use it with your eyes open.