Vast AI review 2026: GPU rental pricing, features, pros and cons

Vast AI (Vast.ai) is a GPU rental marketplace: independent hosts list their graphics cards, and you rent them by the second for AI training, inference or rendering. In November 2026 it's one of the cheapest ways to rent a GPU for AI, and I think it's a great fit for tinkerers and researchers on a budget. The trade-off is reliability. You're renting someone else's machine, and the quality varies from listing to listing.

Nobody at The AI Future ran a hands-on test for this review. We compared Vast.ai using its official docs and live prices, an independent price tracker and 32 recent Trustpilot reviews, as of early November 2026. Where I quote someone's experience, I'll say whose.

Vast AI in brief

Vast.ai connects two groups. Hosts range "from individuals with a single gaming PC to Tier-4 datacenters", and each one sets its own prices. Renters search those offers, pick one and get a Docker container (or, less often, a virtual machine) with exclusive use of the rented GPUs.

The scale is bigger than I'd have guessed. The homepage claims 20,000+ GPUs across 40+ data centers, with prices "set by supply and demand". There's no subscription and no sales call; you add credit and start renting.

Lately Vast has been pitching itself at AI agents too. Its tagline is now "Agent-Ready AI Infrastructure", and the idea is that the same API developers use to launch a gpu server for AI can be called by software that buys its own compute. That's a marketing angle for now, but it's a fair description of how scriptable the whole thing is.

Key features

A live GPU marketplace

Every offer shows the GPU model and count, CPU, RAM, disk, bandwidth, price, location and maximum rental duration. Each also carries two Vast-made numbers, going by the concepts page: a DLPerf score that estimates deep learning speed, and a reliability score based on the machine's uptime history. New machines start at 60% reliability and climb from there.

The range is wide. The pricing page lists 68+ GPU types, from an RTX 3060 up to the B200, and ComputePrices counted 70 models across 186 configurations on November 7, 2026. If you're still fuzzy on why the card matters so much, our explainer on what a GPU does in AI is a quick primer.

Three ways to rent

• On-demand instances have fixed host pricing and high priority, so they aren't interrupted before the contract's maximum duration.
• Reserved instances are on-demand rentals you prepay for up to 50% off, with partial refunds if you cancel early.
• Interruptible instances run on bids, are often 50%+ cheaper, and get paused if someone outbids you or wants the machine on demand.

One quirk from the instance types guide: you can't switch an on-demand rental to interruptible (or back) later. You'd have to start a new instance.

Templates and a model library

Templates are saved launch settings, essentially a wrapper around docker run. Vast ships ready-made ones for PyTorch, TensorFlow and ComfyUI, per the quickstart, and the homepage lists one-click setups for open models such as Qwen3.8 27B and Kimi K3. That saves a lot of fiddling for anyone who doesn't want to build images from scratch.

Serverless endpoints and clusters

Serverless lets you deploy a model behind an endpoint that adds and removes GPU workers as traffic changes. Looking at the serverless billing table, it costs the same per second as normal instances, with no extra fee; idle "inactive" workers stop paying for compute but keep paying for storage. For big training runs, Vast also sells dedicated multi-node clusters with InfiniBand networking.

CLI, SDK and API

Everything in the console can be done from the vastai command line tool, a Python SDK or a REST API. You can even query live offers programmatically with "vastai search offers", which is handy if you want a script to grab the cheapest suitable card at 3am.

Vast AI pricing and plans

There are no plans. You top up credit (the minimum deposit is $5, by card, BitPay or Crypto.com) and pay per second while an instance runs. Hosts set the rates, so prices move daily.

The cheapest on-demand single-GPU listings below come from ComputePrices, which collected them on November 7, 2026, along with its ranking against other providers. Treat them as a snapshot, not a quote.

Prices checked November 2026. ComputePrices says Vast was under the market median on 52 of its 70 GPUs. The pattern's interesting: consumer and older data-center cards are very cheap, while the newest chips (H200, B200) sit around or above the market median. For comparison, Vast's own pricing guide lists RunPod's H100 SXM at $2.99 per GPU-hour and Lambda's at $3.99, checked July 20, 2026.

Costs on top of the hourly rate

The GPU rate isn't the whole bill. Going by the billing docs, storage is charged continuously while an instance exists, even when it's stopped, and bandwidth is charged per byte at each host's own rate. When your balance hits zero, instances stop; if there's no card on file, your data is scheduled for deletion.

Honestly, this is the part I'd read twice. One Trustpilot reviewer, Sora_, complained about bandwidth charges of "$2.5/100gb" after a 20-minute session. And the disk size you choose can't be changed later, so you can't simply grow it mid-project.

Is there a free plan? No. There's no free tier or trial credit on the pricing pages I read; $5 is the entry ticket.

Performance and independent testing

Raw GPU speed isn't really the question. An H100 is an H100 wherever it sits. What changes on Vast is everything around the card: the host's CPU, disk, network and uptime.

That's where the reviews split. Henning N. wrote in December 2025: "Sometimes my instances just stop but that can be ameliorated by aggressive checkpointing." He still gave it four stars. In October 2026, Çetin was much harsher: "Instances went offline at random in the middle of my work, with no warning and no explanation, and I lost hours of unsaved results." Other one-star reviews describe SSH key failures and instances stuck on "running".

A chess-engine developer, Jan Klima, found the price-to-performance ratio good on unverified machines, but warned that "Spent credits are non-refundable." He was charged for a verified instance that went offline overnight.

The Activepieces pricing breakdown makes the obvious point well: on a flaky host, a low hourly rate can cost more than a steadier one, because you pay for hours that produce nothing. I'd guess most bad experiences come from picking the cheapest listing with a low reliability score. Filtering for higher reliability and data-center hosts probably fixes a lot of it, though it costs more.

Scores land in the middle. Vast.ai has a 3.8 TrustScore from 263 reviews on Trustpilot (checked November 7, 2026), with 77% five-star and 16% one-star, a very polarised spread. Support gets surprisingly good marks, though. Trustpilot's AI summary of 87 recent reviews says customers praise a fast, human chat team, and Oscar Laird put it this way: "Extremely helpful and responsive customer support. This surprised me because it is a very cheap/budget gpu rental platform."

Best suited to

• Students, hobbyists and indie developers who want to rent GPU for AI cheaply, for example a 24GB RTX 4090 for Stable Diffusion or LoRA fine-tuning.
• Researchers running batch jobs that checkpoint often, where interruptible pricing pays off.
• Developers who want to script GPU buying through an API or CLI.
• Small AI apps that can tolerate a node dropping, using Serverless to spread work across hosts.

I'd skip it for regulated data or anything where one outage costs real money, unless you pin yourself to verified data-center hosts. If you're weighing renting against owning hardware, our piece on whether on-premise AI is cheaper than cloud AI runs the numbers.

Vast AI alternatives worth considering

• RunPod is a GPU cloud with fixed published rates; Vast's guide lists its H100 SXM at $2.99 per GPU-hour. Pick it if you'd rather not compare hosts.
• Lambda runs its own data centers and lists an H100 SXM at $3.99 per GPU-hour on 8-GPU instances, per the same table. It's better for teams that want one accountable provider.
• Google Cloud Vertex AI is the managed, enterprise route with MLOps tools built in. See our Google Cloud Vertex AI review.

For a broader list, our roundup of AI cloud computing tools and our guide to cloud native AI platforms cover the bigger names.

Verdict

Vast.ai is, in my view, the best value GPU rental around if you're willing to do a little homework. The per-second billing and $5 entry point are hard to argue with, and the prices on older and consumer cards are often the lowest anywhere. But it's a marketplace, and some hosts are flaky; storage and bandwidth fees can surprise you, and spent credit isn't refunded. Use it for experiments, fine-tuning and checkpointed training. Skip it, or stick to verified data-center hosts, for production work you can't afford to lose.

FAQ

What is Vast AI?

Vast AI (Vast.ai) is an online marketplace for renting GPUs. Hosts, from home PC owners to data centers, list hardware and set prices, and you rent it by the second for AI training, inference, image generation or rendering. You launch Docker-based instances through a web console, CLI, Python SDK or API.

Is Vast AI safe and legit?

Yes, it's a real Los Angeles-based company with a claimed Trustpilot profile rated 3.8 from 263 reviews in November 2026. The risk is reliability rather than fraud: hosts vary, instances can drop, and spent credits aren't refundable. For sensitive data, it's probably wise to choose verified data-center machines only.

How much does Vast AI cost?

There are no plans. You prepay credit (minimum $5) and pay per second. On November 7, 2026, ComputePrices tracked its cheapest RTX 4090 at $0.36 per GPU-hour and H100 SXM at $2.27. Storage and bandwidth cost extra, and interruptible rentals are often 50% or more cheaper.

Is Vast AI cheaper than RunPod?

Usually, on consumer and older data-center GPUs. ComputePrices ranked Vast under the market median on 52 of 70 GPUs, and its H100 SXM floor of $2.27 beat the $2.99 RunPod rate in Vast's July 2026 comparison. For the newest chips like the H200, the gap seems to shrink or reverse. Bear in mind those competitor figures are fixed published rates, while Vast's move hourly, so the cheapest Vast listing you see on any given day could be higher or lower than this snapshot, and it may sit on a host with a weaker reliability score.

Is Vast AI an AI operating system?

No. It's GPU infrastructure, not an OS. The "Agent-Ready" pitch just means AI agents can rent compute through its API, while you bring your own models and software.

Can I earn money by hosting GPUs on Vast AI?

Yes, anyone can list a machine as a host and set a price, and Vast has an earnings calculator for hosts. Be careful with expectations, though. On Trustpilot, one host said earnings were "an order of magnitude lower" than the calculator suggested, and a couple of others reported trouble withdrawing their money in 2026.