Vast.ai is a “GPU cloud marketplace” that connects renters with GPUs held by data centers and operators around the world. You can rent GPUs such as the H100, B200, and RTX 4090 by the hour (billed by the second), and prices are not fixed list rates ─ they move in real time with supply and demand. Instead of one company setting the price as with the major clouds, many hosts list their GPUs and you pick from that catalog, so the same GPU can vary widely in price depending on the conditions. It suits anyone who wants to keep costs down on GPU-heavy work: training and fine-tuning machine learning models, running inference servers, or generating images. The service has been available since 2018, and it ships developer-facing APIs, a CLI, and a Python SDK.
Key Features
- Marketplace-driven pricing: Prices are set by supply and demand across 40+ data centers and change in real time. More than 68 GPU types are listed, so you can compare performance, price, location, and host reliability before choosing
- Three instance types: Choose “on-demand” (never interrupted), “interruptible” (may be reclaimed by the host, but far cheaper), or “reserved” (capacity locked in for 1, 3, or 6 months). According to the official site, interruptible instances are 50%+ cheaper than on-demand, and reserved instances go up to 50% off
- Per-second billing: On-demand instances are billed by the second, so short experiments and quick validation runs waste little
- API, CLI, and Python SDK: An OpenAPI specification is published, and the CLI and Python SDK let you automate searching for, launching, and stopping instances ─ useful for scheduling training jobs or automatically renting whenever price conditions are met
- Secure Cloud: For users with heightened security and compliance requirements, there is an option limited to GPUs hosted by certified data center partners. The company holds SOC 2 certification
- Docker-based environments: You launch instances from any Docker image you specify, which makes it easy to bring your existing PyTorch or inference-server setup along as-is
Pricing
Vast.ai has no monthly subscription plans; you pay only for what you use. GPU rates are set by hosts and move with supply and demand, so there is no fixed list price. The differences between instance types are as follows.
| Instance type | Billing | Characteristics |
|---|---|---|
| On-demand | Per-second usage-based | Never interrupted. Suited to production workloads and long training runs |
| Interruptible | Usage-based | May be reclaimed by the host, but 50%+ cheaper than on-demand. Suited to batch training |
| Reserved | 1, 3, or 6 month terms | Capacity is guaranteed, at up to 50% off |
Note that storage and network transfer are billed separately at rates set by each host, on top of the GPU rate (storage keeps accruing even while an instance is stopped ─ it stops only when the instance is deleted). Payment works on prepaid credits that are drawn down as you use the service.
Pricing is as of August 2026. Actual GPU rates change from moment to moment, so check the official pricing page, the pricing calculator, and the listings in the console for the latest figures.
Pros & Cons
✅ Pros
- The same GPU can often be rented considerably cheaper than an equivalent instance on a major cloud
- Per-second billing plus a wide catalog of GPUs means you can use exactly the performance you need, only when you need it
- Interruptible instances can dramatically cut the cost of batch training that can be restarted after a failure
- The CLI, API, and Python SDK let you script the procurement of instances itself
- Because you specify a Docker image, there is little environment setup to redo
⚠️ Cons
- Prices fluctuate, so monthly costs are hard to fix in advance; budgeting needs headroom
- Network speed, disk performance, and stability vary by host, so quality is uneven and vetting listings takes effort
- Interruptible instances can be reclaimed without notice, so checkpointing and interruption-tolerant design are mandatory
- Storage and transfer are billed separately, and stopped instances still incur storage charges, so costs pile up if you leave things running
- Managed training platforms and job management features are thin ─ essentially you are renting a server with a GPU, and orchestration is yours to build
Comparison with Similar Services
| Criteria | Vast.ai | RunPod | Lambda | CoreWeave |
|---|---|---|---|---|
| Delivery model | Decentralized marketplace | GPU cloud (Community/Secure tiers) | GPU cloud on own data centers | Enterprise-focused GPU cloud |
| How prices are set | Supply and demand (host-set) | Published by the provider per type | Published by the provider | Mostly contract-based |
| Tendency toward low cost | Easiest place to chase the lowest rate | Moderate | Moderate | Aimed at large contracts |
| Variance in quality | Varies by host | Relatively consistent | Consistent | Consistent |
| Typical use | Training/inference for individuals and small teams | Inference and training for individuals to teams | Training and research | Large-scale training and production inference |
Even among services that all “rent you a GPU,” the split is clear: Vast.ai when the lowest price matters most, and more managed offerings when uniform environments and hands-off operation matter more.
Who Is It For
- Individuals and small teams looking to reduce GPU costs for training and fine-tuning
- People running large volumes of image or video generation who would rather not buy a high-end GPU
- Anyone who can structure batch work to resume after interruption (they benefit most from interruptible instances)
- Developers comfortable automating infrastructure procurement through a CLI or API
- Researchers and students who need a bigger GPU than their local machine, but only occasionally
Conversely, production systems that require an SLA and uniform quality, or teams that want GPU procurement and operations handled for them, are better served by more managed offerings.
Summary
Vast.ai turns GPUs from something you buy at list price into something you rent by watching the market. The marketplace model makes it easy to chase the lowest rate, but it also hands you more to manage: quality differences between hosts, interruption risk, and separately billed storage. A realistic way to start is to load a small amount of credit, rent a short-lived instance, and actually measure network speed and disk performance before committing to heavier use.