Vast.ai is the largest open marketplace for renting GPUs by the hour, and two numbers from its own listing board explain both the appeal and the risk: an RTX 4090 rents from a floor of $0.14/hr with a median of $0.49/hr. This review walks through how the marketplace is built, what renters get at those prices, where the model bites back, and what the host side pays out.
01. What is Vast.ai, and how is it built?
A marketplace, not a datacenter.
Vast.ai does not own the machines it rents you. It operates a two-sided market: independent hosts — home rigs and small datacenter operators alike — list their GPUs by the hour and set their own prices, while the platform supplies the matching, billing, and container tooling that turns a stranger's machine into a bookable instance. The listing page behaves like an exchange, not a catalog: the same GPU model appears multiple times at different prices because different owners are selling it. When independent owners compete for the same renter, a listing price finds its level the way a price on an exchange does — which is also why the same card never arrives with one number.
A commission on every rental, priced by the platform.
The platform's revenue is a commission on each rental, and the size of that cut is a figure Vast.ai publishes on its own commission page. Hosts set the listing price; the platform takes its share from the transaction rather than charging you a separate booking fee. This review deliberately does not restate the percentages that circulate in third-party posts about host earnings: the only number worth planning around is the one on the platform's own page, and marketplace terms change without notice.
One brand, four product lines.
The site's navigation splits into GPU Cloud, Clusters, Serverless, and Hosting — hourly rentals from the listing board, multi-machine setups for larger jobs, a tier for calling GPUs without managing an instance, and the host program that feeds supply into everything else. For a renter comparing vendors, the first and the last matter most: GPU Cloud is where the listings live, and Hosting is how new machines keep arriving.
02. What do renters actually pay?
The 4090 numbers: a $0.14 floor, a $0.49 median.
At the time of writing, marketplace listings for an RTX 4090 run from a floor of $0.14 per hour to a median of $0.49 per hour — a spread of about 3.5×. That spread is not noise; it is the market's core property. Every host prices its own card, so identical GPU names arrive with different storage, networking, locations, and reliability behind them. The floor is real and bookable, but a budget built on the floor breaks the first time that listing is taken by someone else; a budget built on the median survives ordinary price movement.
Interruptible instances trade security for a lower rate.
Beside ordinary on-demand rentals, Vast.ai lists an interruptible tier: a lower price in exchange for the right to preempt your machine when another renter needs it. An interrupted instance stops and can be resumed later, but the local disk attached to it may not survive the event — which is why interruptible work belongs to jobs that checkpoint to storage you control. There is no fixed discount to quote here, because each host prices its own preemptible listings, and the numbers move with demand.
Setup is a template away.
Vast.ai maintains official preinstalled templates for common workloads, including a ComfyUI image with a dedicated Wan 2.2 variant — ComfyUI being the node-based interface most image and video workflows are built on, and Wan 2.2 being the current open video-model family. Booking a template means the environment boots ready to generate instead of starting from a CUDA install, and on a machine billed by the hour, setup time is pure cost.
| Listing property | What you see | What it means for your job |
|---|---|---|
| RTX 4090 price | $0.14/hr floor, $0.49/hr median | Budget on the median; treat the floor as opportunistic |
| On-demand rental | Host-set hourly rate | The machine stays yours until you stop it |
| Interruptible rental | Cheaper host-set rate | Can be preempted; resumes after interruption, keep data off the local disk |
| Hardware source | Third-party hosts, individual listings | Check the host's rating before booking |
| Environment | Official templates, including ComfyUI with Wan 2.2 | Setup cost is close to zero for common workloads |
03. What is Vast.ai good at?
A floor that competition sets.
Because every host prices its own hardware, the $0.14/hr floor for an RTX 4090 comes from owners bidding for renters, not from a rate card one product manager wrote. A first-party cloud reprices its catalog a few times a year; a marketplace reprices continuously, and a renter flexible about timing catches the dips. Price-sensitive users are the reason this platform exists, and the floor is where that shows.
A card list that spans consumer to datacenter.
The listings run from consumer cards such as the RTX 3090, RTX 4090, and RTX 5090 to datacenter hardware such as the H100, so one account covers a 24GB image-generation box tonight and an 80GB training machine next week. That breadth matters for mixed workloads: the card that renders video and the card that fine-tunes a model sit in the same search filter instead of in two vendor contracts.
Supply that keeps arriving.
Supply comes from independent hosts adding machines continuously rather than from a single warehouse, which is why specific cards appear and disappear through the day and why the marketplace can absorb demand spikes that a fixed fleet would throttle. The same depth has a flip side — it is the reason hardware quality differs between listings, which is the next section's subject.
04. What does the marketplace model cost you?
No fixed performance promise.
A listing is an offer from an individual host, so Vast.ai does not stand behind fixed performance, uptime, or service-level commitments the way a first-party cloud does. Two listings with the same GPU name can deliver different sustained clock speeds, different cooling headroom, and different storage latency. Treat the platform's reliability signals — ratings, uptime statistics, and benchmark figures shown per machine — as the substitute for a contract, and read them before booking rather than after.
Interruption is the price of the cheap tier.
The interruptible discount is not a promotion; it is compensation for preemption risk. Renters who book interruptible instances accept that a busier period can reclaim the machine mid-job. The failure mode to design for is simple: the instance stops, the local disk may not come back, and the job resumes from wherever it last checkpointed. Workloads that cannot tolerate that belong on on-demand listings, paid at the median.
Third-party hardware needs vetting.
Quality control on a marketplace is distributed to the renter. Machines range from maintained datacenter rows to a gaming rig in a warm room, and the same GPU model can behave differently behind them. The practical routine: filter by the host's rating and the machine's verified benchmarks, start with a short rental before committing hours, and read an unfamiliar host's floor price as compensation for an unproven machine rather than as a bargain.
05. What does the host side look like?
Income equals list price minus commission, paid as cash.
For machine owners, the deal reads differently than it does for renters. A host sets the listing price for its card, the platform takes its commission from each rental, and the payout arrives as cash — a different shape than the gift-card credit that Salad pays its contributors or the token rewards io.net layers on for its supply network. The structure is straightforward; the outcome is not guaranteed, because earnings depend on the price you set, the hours the card actually rents, and demand for that card class, and none of those come as promises from the platform.
The earnings calculator models scenarios, not outcomes.
Vast.ai's host program includes an earnings calculator that projects what a card could make under assumptions you choose — your price, your card class, your availability. Its output is a planning input for deciding whether to list a machine, not a commitment the platform makes to you. The host documentation describes the underlying arrangement plainly: you list a machine you already own, it rents by the hour when you are not using it, and it goes back to serving you between rentals.
06. Who fits Vast.ai — and who doesn't?
A three-line fit test.
Vast.ai fits you if you are price-sensitive, can absorb an interruption on the cheap tier, and will spend a few minutes vetting hosts before booking. It fits poorly if your workload needs contractual uptime, fixed performance, or a machine nobody else can reclaim — the properties a first-party cloud sells and a marketplace structurally cannot. Between those poles, most experimental workloads — image generation, video renders, model evals, burst capacity beside a small local rig — land closer to the marketplace end, which is why the floor-and-median spread is a feature to plan around rather than a defect to complain about.
SpareHive is a waitlist for task-based image and video jobs: a preinstalled environment, a prompt or job in, files back, no SSH, and no multi-gigabyte weight download over a slow link. Per-second billing and a $5 minimum top-up are the plan. The marketplace is not taking rental jobs yet. Join the waitlist at sparehive.com.
FAQ
Is the $0.14/hr RTX 4090 listing the price to plan on?
No. Plan on the $0.49/hr median: the floor is one host's listing, it disappears when that machine is taken, and a budget built on it breaks at the first rebooking. Use the floor for opportunistic short jobs and the median for anything with a deadline.
Does Vast.ai guarantee a listing will perform like a reference card?
No. Listings come from independent hosts, the platform does not stand behind fixed performance or uptime, and two machines with the same GPU name can behave differently. Filter by host rating and verified benchmarks, and rent briefly before committing long hours.
How large is Vast.ai's host commission?
No fixed figure belongs in this answer. Vast.ai publishes its commission rate on its own pages, and that page — not any blog post, including this one — is the figure to trust, because marketplace terms change without notice.
Can an interruptible instance cost me data?
Yes. A preempted interruptible machine stops without warning and its local disk may not survive the event, so checkpoints and outputs belong on storage you control. If a job cannot resume from a checkpoint, it does not belong on the interruptible tier.