Vast AI review: quick verdict
Vast.ai fits teams that can compare individual offers, checkpoint important work, and account for storage and network charges.[source][source][source] It is a poor fit when the workload cannot recover from interruption or procurement requires uniform controls across every host.[source][source] Sensitive projects need a separate security review before deployment.[source]
Key takeaways
- Vast.ai is a marketplace. Hosts set offer terms, so renters must assess each machine and contract rather than assume one uniform service.[source]
- The complete bill can include compute, storage, and bandwidth. Stopped instances can remain online and accrue storage, while Vast's detailed billing page says offline instances accrue neither active rental nor storage.[source][source]
- Vast documents on-demand, reserved, and interruptible rentals. Interruptible work needs external checkpoints and saved outputs.[source]
- Verification is a platform status, not an uptime promise, and Vast says provider security varies.[source][source]
- Serverless is a separate endpoint and worker model for variable request demand, not simply another name for a rented instance.[source]
Vast.ai trades a standardized fleet for offer-level flexibility. Teams that are comfortable inspecting offers and building recovery into their jobs may accept that trade. Teams that want the provider to make most infrastructure choices should compare the broader GPU cloud provider options before depositing funds.
How Vast.ai works
Vast.ai connects renters with GPU hosts that range from individuals to datacenters. A host publishes offers for a machine, and the renter chooses an offer and creates a rental contract. The marketplace displays details for each offer:[source]
- GPU
- CPU
- Memory
- Storage
- Bandwidth
- Price
- Maximum duration
- Location
- DLPerf
- Reliability score
That structure differs from renting from a single, centrally operated fleet. Two offers for the same GPU model can have different surrounding hardware, network terms, locations, rental limits, and host status.[source] The offer is the unit to evaluate. A familiar GPU name alone does not establish that two machines are interchangeable.
Vast documents an instance as generally being a Docker container, with a smaller set of virtual machine offers also available.[source] Templates define the software image and launch settings, while the selected offer supplies the underlying resources.[source] This can make familiar container workflows possible, but it does not remove the need to inspect storage persistence, ports, environment settings, and the host information attached to the offer.
One static price cannot represent the Vast.ai marketplace. Hosts set their own prices, and supply and demand can change the listings.[source] The public pricing page is intended to show live platform rates, but the source available for this review did not expose those offer values.[source] Quoting an old hourly rate would be misleading. Confirm the selected offer in Vast.ai immediately before starting a rental.
Vast.ai pricing and total cost
The GPU rate is only one part of a Vast.ai cost estimate. Vast documents separate compute, storage, and bandwidth charges, with the actual terms depending on the offer.[source][source] A useful budget therefore starts with the job's runtime and data movement, not with a screenshot of one marketplace listing.
| Cost area | What to inspect before renting | Why it affects the decision |
|---|---|---|
| Compute | Rental type, host-set GPU rate, and expected active time | The rate can change across offers and rental models.[source] |
| Storage | Allocated space, current online or offline state, expected retention, and cleanup plan | Storage is billed while an instance is online, including when manually stopped, but not while it is offline.[source][source] |
| Bandwidth | Upload terms, download terms, and output size | Data transfer can be charged separately from compute.[source][source] |
| Recovery | Checkpoint destination and restart process | An interrupted job may need a new rental and restored state.[source] |
Vast documents how each component is billed:[source]
- Active rental: by the second
- Online storage: by allocated size
- Upload and download bandwidth: by the byte
The detailed billing page defines online as every state other than offline. It says a manually stopped instance can still incur storage charges, while an offline instance incurs neither active rental nor storage charges.[source] The broader pricing page also warns that stopped instances accrue storage.[source] Check the state shown in the console and delete resources that are no longer needed.
Account balance and data retention also belong in the runbook. Vast's billing documentation distinguishes active and offline instances and explains how account funds affect resources.[source] Before moving valuable data onto a rental, check the current billing rules in the console, keep outputs outside the instance, and define what should happen if funding or the rental term ends. The GPU price lookup tool can help narrow models, but the final cost check belongs on the live Vast offer.
On-demand, reserved, and interruptible instances
Vast documents three rental types: on-demand, reserved, and interruptible.[source] Choosing among them is an operational decision, not simply a hunt for a smaller number.
On-demand rentals
An on-demand offer uses a fixed price set by the host and receives high marketplace priority. The offer also displays a maximum rental duration.[source] High priority means it is not subject to the marketplace preemption rules used for interruptible rentals. It does not promise that hardware, networking, funding, or every other dependency can never fail.
Maximum duration matters for long jobs and retained data. Vast says an expired instance may be deleted 48 hours after expiration.[source] Treat the displayed duration as a hard planning input. Export checkpoints and results before the deadline rather than relying on a later recovery window.
Reserved rentals
A reservation starts by prepaying an existing on-demand instance. The prepaid credit is locked to that instance, and the reservation cannot move to another host.[source] This model may fit steady demand after a team has already evaluated the machine and expects to keep using it.
Prepayment raises the cost of a poor selection. Confirm the GPU, surrounding resources, host terms, maximum duration, and cancellation conditions before reserving. A reservation should follow a workload forecast and an offer review. It should not be the first step taken merely because a promotional discount is displayed.
Interruptible rentals
An interruptible instance can be paused when another renter outbids it or when on-demand demand takes priority. Vast tells renters to checkpoint work and save important outputs outside the instance.[source] This rental type fits work that can resume safely, such as a batch process with tested checkpoint restoration.
Do not choose interruptible capacity for a job that writes its only copy of state to the rental. Before launch, test the restore command and decide how much repeated work is acceptable.
How to evaluate a Vast.ai offer
Start with workload constraints. Write down the GPU memory requirement, GPU count, host memory, storage need, output size, expected runtime, and acceptable locations. Then compare only offers that meet those constraints. This prevents an attractive GPU rate from hiding an unsuitable maximum duration or expensive data path.
Next, read the whole offer. Vast exposes machine and contract details that include resources, network information, maximum duration, location, performance information, and reliability information.[source] DLPerf and the reliability score are Vast platform fields. They are not GPU Picks measurements, and neither field should be converted into an expected completion time or uptime percentage.
Machine verification needs similar restraint. Vast describes verification as an automated platform process. A machine can remain unverified because it is new or still being evaluated, while a Verified machine has passed Vast's documented checks and threshold.[source] Verified does not mean that failure is impossible. Unverified does not prove that a host is malicious. Use the status as one input alongside rental terms, workload sensitivity, and recovery design.
A practical offer review asks:
- Does the machine meet the actual memory and storage requirements?
- Does the maximum duration leave enough time for the job and export?
- Can the job recover if this rental ends earlier than planned?
- Where will checkpoints and final outputs live?
- What will compute, retained storage, and data transfer cost together?
- Is the host category acceptable for the data and procurement requirements?
Save the offer details used for the decision. Marketplace listings can change, so a team needs a record of what it approved and a fresh check before each material run.
Security boundaries and sensitive workloads
Vast says client workloads run in unprivileged Docker containers, but its security guidance also says provider security varies.[source] Container isolation is relevant, yet it does not make every host or facility equivalent. A threat model should account for the host, workload secrets, data sensitivity, network exposure, and the renter's own configuration.
For sensitive work, Vast recommends Secure Cloud along with encryption and external key management.[source] That is provider guidance, not a substitute for legal, security, or compliance review. Keep decryption keys outside the rented host, limit credentials to the permissions the job needs, and avoid placing the only copy of sensitive data on the instance.
Vast publicly states that it has completed SOC 2 Type 2 and that its report is available under NDA.[source] GPU Picks has not inspected that report, and the statement does not establish that every marketplace host has the same controls. The Security FAQ says Secure Cloud providers have ISO 27001 certification and Tier 3 or Tier 4 datacenter standards.[source] The compliance page instead says partner certifications may include ISO 27001 or SOC 2 and are encouraged but not strictly required.[source] It also calls Verified Hosts manually tested, while the verification documentation describes an automated process.[source][source] These official descriptions conflict. Teams with formal requirements should ask Vast which controls, certifications, and verification process apply to the selected host.
Vast Serverless
Vast Serverless uses endpoints that scale worker instances for incoming work. Vast positions the model for variable or unpredictable demand rather than requiring a renter to keep one instance active continuously.[source][source] This is a different operating model from manually selecting and maintaining a rental.
Vast says a new Serverless endpoint goes through an evaluation period that benchmarks the workload across GPU classes.[source] The public source used here does not establish a universal evaluation duration, request latency, startup time, throughput, or production capacity. A team considering Serverless should evaluate its own container, request pattern, scaling settings, error handling, and billing behavior before routing production traffic.
Serverless can be relevant when requests arrive unevenly and workers should follow demand. A persistent instance may be simpler for interactive development, a continuously busy service, or a job that needs durable local state. Compare both models using the same workload and complete cost assumptions.
Who should use Vast.ai
Vast.ai is a plausible candidate for researchers, engineering teams, and startups comparing GPU cloud options that run resumable jobs and can review offer-level terms. Someone still needs to own selection, budget checks, data handling, and recovery.
Who should skip Vast.ai
Skip Vast.ai, or narrow the search to an approved host category, when a marketplace contract creates unacceptable uncertainty or nobody can review each offer. Sensitive workloads also need host-level evidence for required certifications, facilities, and controls. A small team may prefer a standardized service if offer review and recovery work would consume the engineering time saved on infrastructure cost.
Vast.ai alternatives
Read the RunPod review when comparing a different product model. Read the Lambda GPU Cloud review when comparing a more standardized instance-selection path. These links support further research; neither is an automatic winner. Compare current terms, GPU fit, storage, network policy, capacity access, and recovery requirements. The GPU cloud reviews index provides the available provider research in one place.
Pros and cons
Pros
- Host-set offers provide different combinations of machine terms and rental models.[source]
- Vast documents on-demand, reserved, and interruptible contracts.[source]
- Offer details expose duration, resource, network, location, and platform-status fields.[source]
- Serverless provides a separate endpoint and worker option for variable request demand.[source]
Cons
- A renter must evaluate the complete offer rather than rely on the GPU model or compute rate alone.[source][source]
- Online storage remains billable when an instance is manually stopped, while offline instances incur no storage charge.[source]
- Interruptible work needs external checkpoints and a tested restore process.[source]
- Sensitive workloads require host-level security and procurement review because provider controls vary.[source]
Decision limitations and methodology
This review is based on Vast.ai's public pricing, product, billing, security, and compliance documentation. GPU Picks did not rent an instance, inspect a host, review Vast's SOC 2 report, sample capacity, or collect workload performance measurements. Read the full GPU Picks research methodology before using the editorial conclusions.
Marketplace prices, availability, maximum duration, and host status can change. The live rate cards were not available in the public source captured for this review, so no hourly values appear here. Recheck the live offer and applicable contract immediately before spending money or moving data.
Frequently asked questions
Is Vast.ai safe to use?
Vast.ai can be considered only after matching the host category and controls to the workload. Vast says provider security varies and recommends Secure Cloud, encryption, and external key management for sensitive work.[source] Do not assume container isolation or a platform compliance statement settles host-level procurement requirements.
How does Vast.ai pricing work?
Hosts set marketplace prices, so offers can differ and move with supply and demand.[source] The complete cost can include compute, storage, and bandwidth.[source] Review the live offer and estimate data transfer plus storage for every online state, including a manual stop. Vast's detailed billing page says offline instances incur no storage charge.[source]
What is the difference between on-demand and interruptible instances?
On-demand rentals use a fixed host-set price and receive high marketplace priority. Interruptible rentals can be paused when outbid or displaced by on-demand demand.[source] Choose interruptible only when the job saves checkpoints and outputs externally and the team has verified that restoration works.
Does Vast.ai charge when an instance is stopped?
A stopped instance can remain online, and Vast says storage charges continue in that state.[source][source] The detailed billing page says an offline instance incurs neither active rental nor storage charges.[source] Check the console state rather than treating every non-running instance the same, and delete resources that are no longer needed.
What does Verified mean on Vast.ai?
Verified is a Vast platform status produced through its automated verification process. Vast says the machine has passed documented checks and a reliability threshold.[source] It is useful evidence about the platform's evaluation, but it is not an uptime guarantee or proof that every future job will complete.
Does Vast.ai offer serverless GPUs?
Yes. Vast Serverless creates endpoints that scale worker instances for incoming demand, and new endpoints go through an evaluation period across GPU classes.[source] Evaluate the workload's own request behavior, scaling settings, errors, and cost because the documentation does not establish one universal latency or startup result.