Vast.ai is a marketplace where independent hosts and data centers rent out GPUs, with on-demand, interruptible and reserved instances for AI training, inference, fine-tuning and rendering.
Used this tool? Rate it
Used this tool? Rate it
Vast.ai is a GPU rental marketplace that connects people who need compute with hosts who rent out their hardware. The company describes the service as a cloud computing, matchmaking, and aggregation service focused on lowering the price of compute-intensive workloads.
The model is two-sided. Hosts run Vast.ai's management software, list their machines and set their own prices; renters search the offers, rent a machine, and then run commands or open SSH sessions with a few clicks. On the renter side, the GPU Cloud product offers on-demand instances across 40+ data centers and 20,000+ GPUs, deployable through the CLI, SDK or API. The company also positions itself as infrastructure that AI agents can use to procure their own compute, with API-native provisioning, real-time pricing and per-second billing.
The most important fact for a first-time renter is structural: Vast.ai states that it is only a marketplace and does not manage or provide the hardware. Price, uptime, security and network quality therefore depend on the host you pick.
Jake Cannell and Christian Horne incorporated Vast.ai as a Delaware C Corporation on June 28, 2016.
| Rental type | Priority | How it is priced |
|---|---|---|
| On-demand | High | Fixed pricing set by the host, with guaranteed resources |
| Reserved | High | An on-demand rental converted to a discounted rate through pre-payment |
| Interruptible | Low | Lowest cost, bid-based, and may be paused |
Interruptible instances may be paused if they are outbid or if someone requests the machine on-demand. Data is preserved while paused, but the instance is not functional until it regains priority and resumes automatically. Once rented, an instance's type cannot be changed; the only allowed switch is converting on-demand to reserved for a discount.
Every offer card also carries quality signals. The reliability score measures a machine's historical uptime and health, and every machine starts at 60%. DLPerf is Vast.ai's own scoring function for estimating performance on typical deep learning tasks. Offer cards also show specs, price and the maximum rental duration; renting creates a contract with the host based on the offer's current terms.
Instances start from templates. Pre-built setups include PyTorch, TensorFlow and ComfyUI, and you can build custom templates.
Beyond Docker containers, Vast.ai offers full virtual machines for workloads that need init managers, nested containerization or process tracing. The trade-offs are concrete: slower creation and boot times, higher disk overhead, fewer available machines and preconfigured templates, and SSH as the only launch mode.
Serverless deploys models as endpoints with automatic benchmarking across GPU types and can autoscale to zero. It charges per second at the same price as regular GPU instances. Clusters are dedicated multi-node GPU clusters with InfiniBand networking, aimed at large-scale training.
A reserved instance is not rented directly. You first rent on-demand, then convert it through the discount badge on the instance card and choose a pre-paid period.
The documentation ties each rental type to a kind of workload:
The homepage frames the platform as covering everything from training to inference and from fine-tuning to rendering.
Two vendor-published case studies show the kind of buyer Vast.ai targets. Creatix Technology, an AI app company, is said to have scaled to 200K daily users while cutting infrastructure costs by over 60%. PAICON, a global oncology data platform, used the GPU cloud to iterate on its Athena model while significantly reducing research-phase training costs.
Good fit:
Weaker fit:
The pages are not fully aligned. The general FAQ says Vast.ai provides Linux Docker instances only and no remote desktop interfaces, while the VM guide describes full Linux VMs and lists a template described as an Ubuntu desktop with a graphical environment. Both pages agree that the platform is Linux-only.
Vast.ai pricing is not a fixed rate card. The platform uses a marketplace model where hosts set their own prices, so rates differ by host and move with demand. Every rental bill combines three charges:
| Charge | Unit | When it accrues |
|---|---|---|
| GPU (active rental) | $/hr, billed per second | Every second the instance is active or connected |
| Storage | $/GB/hr | Every second the instance exists and is online, including while stopped |
| Bandwidth | $/TB | Every byte sent or received, in any instance state |
Bandwidth is easy to underestimate: upload and download costs are not shown in the $/hr or $/day breakdowns because they are charged by usage rather than at a constant rate.
Accounts run on prepaid credits, paid by card through Stripe or by crypto through BitPay and Crypto.com. The minimum deposit is $5. Autobilling can top up the balance from a saved credit card when it runs low. As a reference point for marketplace rates, the GPU Cloud page advertises H100 access for as little as $0.90/hour as captured on October 4, 2026.
The billing page and the terms describe refunds in three different ways, and none refunds spent credits:
Marketplace prices can also move for reasons unrelated to the renter. A June 2026 IT media report on a research preprint attributes a roughly 38% jump in budget GPU rental prices on Vast.ai to demand from a crypto-mining network, with utilization rising from 57% to 94%.
The usual comparison is with hyperscale clouds. Vast.ai's GPU Cloud page claims savings of up to 80% compared to AWS, Azure, or GCP, while its general FAQ puts the reduction at about 3x to 5x; both are vendor claims.
Among GPU-specific clouds, Runpod is the closest structural match: its Pod documentation compares a Secure Cloud option with a Community Cloud option. Runpod also says it is no longer accepting new hosts for Community Cloud, although existing Community Cloud resources remain available. Salad goes further toward consumer hardware: its Community Cloud is a distributed pool of idle consumer GPUs with capacity across 190+ countries.
User feedback points in both directions. As captured on October 4, 2026, a third-party review platform showed a 3.9 score from 257 reviews, with 78% five-star and 15% one-star ratings. Its AI-generated summary of 82 reviews reports praise for support and pricing, alongside complaints about occasional performance drops, unexpected internet bandwidth charges and slow responses during technical issues.
No. Vast.ai says you are not charged while an instance shows "Loading"; GPU charges start once it is active.
No. Billing is per second, so deleting a server after 10 minutes is charged as 10 minutes of usage, plus any bandwidth used.
No. Vast.ai states that instances showing in your account are never free, even when the balance is negative or zero, because storage keeps accruing until the instance is deleted.
