Price checkPer GPU
Interruptible$0.1/GPU/moOn-Demand$0.3/GPU/moReserved$0.2/GPU/mo
See all 3 plans

All Vast.ai Plans & Pricing

Plan Monthly Annual Best For
View all features by plan (compare side-by-side)

Interruptible

  • 50%+ cheaper than on-demand rates
  • Per-second billing, no minimum hours
  • Preemptible — may be reclaimed by host
  • Ideal for fault-tolerant batch workloads
  • Checkpoint and resume supported
  • Lowest entry: TESLA V100 from $0.02/hr

On-Demand

  • Guaranteed uptime, no interruptions
  • Per-second billing, no minimum hours
  • Spin up/down anytime
  • RTX 4090 from $0.30/hr
  • RTX 5090 from $0.37/hr
  • A100 80GB from $0.73/hr
  • H100 SXM from $1.87/hr
  • H200 from $3.44/hr, B200 from $3.85/hr

Reserved

  • Up to 50% discount vs on-demand
  • 1, 3, or 6-month commitment terms
  • Guaranteed capacity for commitment period
  • Volume discounts available
  • Contact Sales required
See Vast.ai Plans
Compare Vast.ai with alternativesAdjust seats, lock a tier, add up to 2 more products side-by-side. Shareable URL.
Cost calculator

What does Vast.ai actually cost you?

Drag the slider. Pick a tier. Your projected spend updates as you go.

Tier
Billing
Your projected cost$2.5per month · $0.1/unit × 25units
Year 1 license$3012 months at this rate
At a glance

List price by tier (annualized, per unit)

Per-unit list price across Vast.ai's plans, annualized. Custom-priced tiers show a hatched bar.

Interruptible$1.2/yr
On-Demand$3.6/yr
Reserved$2.4/yr
Quick Answer
Last verified:
Medium confidence

Vast.ai costs $0.02 to $3.85 per GPU/hour as of July 2026, with 3 plans available. Plans: Interruptible at $0.1/GPU/hour, On-Demand at $0.3/GPU/hour, and Reserved at $0.2/GPU/hour. Pricing depends on your chosen tier, contract length, and negotiated discounts.

Use the interactive pricing calculator to estimate your exact cost based on team size and requirements.

  • Free tier: No free tier available

Vast.ai offers 3 pricing tiers: Interruptible, On-Demand, Reserved. Paid plans include Interruptible at $0.1/user/month, On-Demand at $0.3/user/month, Reserved at $0.2/user/month. The On-Demand plan is production workloads requiring guaranteed uptime.

Compared to other ai/gpu cloud compute software, Vast.ai is positioned at the budget-friendly price point.

  • 8 documented hidden costs beyond list price
Last verified: | Sources: 2 pricing pages and reviews

How much does Vast.ai cost?

Vast.ai pricing ranges from $0.02 to $3.85/GPU/hour across 3 plans. Plans include Interruptible at $0.10/GPU/hour, On-Demand at $0.30/GPU/hour, Reserved at $0.20/GPU/hour.

Vast.ai Pricing Overview

Vast.ai has 3 pricing plans ranging from $0.02 to $3.85/GPU/hour. The Interruptible plan costs $0.10/GPU/hour, best for fault-tolerant batch training and inference. The On-Demand plan costs $0.30/GPU/hour, best for production workloads requiring guaranteed uptime. The Reserved plan costs $0.20/GPU/hour, best for steady, long-running workloads with predictable demand.

There are at least 8 documented hidden costs beyond Vast.ai's list price, including implementation, training, and add-on fees.

This pricing was last verified in May 5, 2026 from 2 independent sources.

See Vast.ai Plans

Vast.ai pricing is usage-based, charged per hour with no free tier or trial. The Interruptible plan starts at approximately $0.10/hour for preemptible GPU instances, the Reserved plan at approximately $0.20/hour for committed capacity, and the On-Demand plan at approximately $0.30/hour for reliable, non-preemptible access. Actual GPU costs vary by hardware type — H100s have been cited at roughly $2.95/hour on the platform — and bandwidth usage is billed separately on top of the hourly rate.

How Vast.ai Pricing Compares

Compare Vast.ai pricing against top alternatives in AI/GPU Cloud Compute.

Hyperbolic $0.3-$3.2/GPU/hour Compare → RunPod $0.34-$3.49/GPU/hour Compare → Lambda $0.69-$6.99/GPU/hour Compare →

Compare Vast.ai vs Alternatives

Before committing to Vast.ai, compare pricing with these 3 alternatives in the same category.

VSHyperbolic
From $0.3/GPU/hour

Full comparison
VSRunPod
From $0.34/GPU/hour

Batch workloads, training runs, and cost-sensitive inference that can tolerate interruptions

Full comparison
VSLambda
From $0.69/GPU/hour

Development, fine-tuning, and small inference jobs

Full comparison

All Vast.ai alternatives & migration guides

What Companies Actually Pay for Vast.ai

Review scores
Trustpilot 1.9out of 5 (25)
Third-party review aggregates, as of Jul 2026
Top pricing complaints
Instances go offline without warning with no billing pause or refundCharged for non-functional or inaccessible instancesBandwidth/egress fees are unexpected and poorly disclosedCustomer support unresponsive or unhelpful, often just copying documentation

Vast.ai Year 1 Total Cost by Company Size

Real deployment costs including licenses, implementation, training, and admin — not just the sticker price.

Serverless AI Inference (Single Model) $0 Year 1 total
$18/hour
Total $0

Running open-weight AI models such as DeepSeek or GLM via Vast.ai's serverless per-minute inference feature, with no GPU instance management required

Single H100 — Short AI Training Run $2 Year 1 total
~$70.80/day
Total $2

Renting one H100 GPU on On-Demand at market rates for a multi-day training job

Interruptible GPU — Budget Experimentation From ~$0.10/hour (Interruptible tier starting price) Year 1 total
Interruptible tier starting price
Total From ~$0.10/hour (Interruptible tier starting price)

Using the Interruptible tier for non-critical experiments or batch workloads that can tolerate preemption

On-Demand Reserved GPU — Production Workload From ~$0.20/hour (Reserved tier starting price) Year 1 total
Reserved tier starting price
Total From ~$0.20/hour (Reserved tier starting price)

Reserved instance for stable, continuous workloads requiring guaranteed availability

hn (2025-12-28)

How Vast.ai Pricing Compares

Software Starting Price Top Price
Vast.ai $0.02/GPU/hour $3.85/GPU/hour
CoreWeave $6.27/instance/hour $68.8/instance/hour
Hyperbolic $0.16/GPU/hour $3.5/GPU/hour
Lambda $0.69/GPU/hour $6.99/GPU/hour
Modal Free $250/GPU/hour
Paperspace $0.45/GPU/hour $5.95/GPU/hour

Detailed pricing comparisons:

Browse all AI/GPU Cloud Compute pricing →

8 Vast.ai Hidden Costs Beyond the List Price

Beyond the listed price, Vast.ai has at least 8 documented hidden costs that can significantly increase total cost of ownership.

Watch for 8 hidden costs
  • Serverless Inference Pricing Premium vs. API Alternatives $18/hr
    high 1 source
    Hacker News "vast offers serverless per minute AI models so checking it for something like https://vast.ai/model/deepseek-v3.2-exp or even glm 3.6 basically every of these turns out to be $.30 cents/minute or 18$ per hour"
  • Interruptible Instance Interruption Loss 10-30% of license costs
    medium 1 source
    Hacker News "Your pricing is around 50-100% higher than vast.ai, a relatively mature competitor in this space who offers the same GPUs and more with beefier rigs."
  • Bandwidth / Egress Charges $2.50/100GB
    high 1 source
    Trustpilot "On top of the price/h of your specific server you will be also charged ungodly amounts for ***internet bandwidth*** in my case it was $2.5/100gb. keep in mind that i only used the instance for 20 minutes."
  • Billing Continues After Instance Becomes Inaccessible 5-25% of total spend
    critical 4 sources
    Trustpilot "I rented a high-end GPU server (RTX 3090 Ti) on Vast.ai with strong specs and a price around $0.143/hour (~0.54 SAR/hour). However, the service is practically unusable."
    Trustpilot "Vast.ai continued to charge my account, even though I had zero access to the server. This is unacceptable for a service that claims to be reliable and user-friendly."
    Trustpilot "I started a verified VM instance, which then was running for almost a day. Checking the progress on the next day, I found that the instance was turned offline without any warning somewhere along the way, meaning that I cannot access my data."
    Trustpilot "if you don't stop the instance before destroying THEY WILL charge you further as they still consider it to be running."
  • Charges for Failed or Non-Functional Instances 5-15% of license costs
    high 3 sources
    Trustpilot "I was charged $1.70 for ~20 minutes of completely non-functional service. Support claimed the charge was only $1.20 for a model download, which is an unacceptable excuse for downtime and unfair billing."
    Trustpilot "Many of the lowest-priced instances are full of bugs and issues. In the end, those supposedly cheap options don’t function properly, and the ones that do work are the more expensive ones."
    Trustpilot "Good cloud provider, but a lot of bad / non working machines in there. Even if you report a non working instance is uses your credits and never gets removed."
  • GPU Performance Degradation Without Price Adjustment 10-25% effective cost increase
    high 2 sources
    Trustpilot "I discovered the host had silently reduced the GPU performance by 22% with no notice and no price adjustment — I was paying full price for a degraded machine."
    Trustpilot "Rented instances removed GPUs while still charging the full price. They restarted my instances without warning on multiple occasions."
  • Trial-and-Error Cost from Unreliable Instances 5-20% of total spend
    medium 1 source
    Trustpilot "I had to rent five different instances before finding one that actually worked, which was a waste of both time and money (not a huge amount). It’s honestly hair-pulling frustrating."
  • Data Loss Risk from Sudden Instance Termination 5-15% of license costs
    critical 2 sources
    Trustpilot "your instance gpus can get scheduled from another guy and you have to copy data which costs bandwidth to another instance just so you can open another even if its targeted as "on demand""
    Trustpilot "My data was stuck on a machine I couldn’t access, my time was wasted for more than 24 hours, and the overall stress caused by their unresponsive support was simply terrible."
Tip

Ask your Vast.ai sales rep about these costs upfront. Getting them in writing before signing can save you from surprise charges later.

Full hidden costs breakdown →

Intelligence sourced from 2 independent sources
Hacker News Tech community Trustpilot Consumer reviews
Key claims include inline source attribution. Data verified against multiple independent sources. 24 source citations total.

Vast.ai Contract Terms

Vast.ai contracts do not auto-renew. Changes require Pay-as-you-go, no formal renewal cycle — instances are billed per hour until stopped or destroyed. These terms are sourced from verified buyer experiences.

Contract Terms
Auto-Renewal No
Cancellation Notice Pay-as-you-go, no formal renewal cycle — instances are billed per hour until stopped or destroyed
Minimum Commitment None for Interruptible and On-Demand; Reserved instances may carry a minimum commitment period
Mid-Term Downgrade Allowed
Payment Terms Prepaid credit balance; usage is deducted continuously at the hourly rate
Price Escalation No published price escalation schedule; spot market pricing on Interruptible tier means effective rates fluctuate with supply/demand
Note

Users can switch instance types when launching new instances; existing instances cannot be downgraded in-flight

Based on 1 verified source

How to Negotiate Vast.ai Pricing

Vast.ai contracts are negotiable. These 8 tactics are sourced from real buyer experiences and procurement specialists.

Negotiation Playbook 8 tactics
Use Interruptible Tier for Checkpoint-Tolerant Workloads high success

The Interruptible tier at $0.10/hr is roughly 66% cheaper than On-Demand at $0.29/hr. Design training pipelines and batch workloads to checkpoint frequently so that preemption results in minimal wasted compute. This is the single highest-leverage cost reduction available on Vast.ai.

Current tier data analysis
Compare Vast.ai Serverless vs. API Providers Before Choosing high success

For inference-only workloads, Vast.ai serverless pricing ($0.30/min, $18/hr) is substantially more expensive than token-based API providers. Evaluate whether you actually need raw GPU access or if an API provider would meet your needs at a fraction of the cost.

hn (2025-12-28)
Choose On-Demand Over Interruptible for Critical Work high success

The Interruptible tier at ~$0.10/hour is significantly cheaper but can be preempted at any time. For workloads where data loss or interruption has real cost (lost compute time, re-running jobs), the On-Demand tier at ~$0.30/hour is often the better economic choice despite the higher headline price.

Trustpilot user reviews
Always Destroy (Not Just Stop) Instances When Done high success

Stopping an instance without destroying it may continue billing on some configurations. Users report being charged after thinking a session was ended. Always explicitly destroy the instance and confirm it no longer appears in your active instances list.

Trustpilot user reviews
Sort Marketplace Listings by Price medium success

Vast.ai is a competitive GPU marketplace where individual hosts set their own prices. Sorting available instances by hourly rate often surfaces options well below the tier minimums listed, especially for older or less-popular GPU models that still handle many ML workloads adequately.

hn (2023-06-15)
Filter for Verified Hosts Only medium success

Unverified instances are cheaper but carry much higher rates of SSH failure, broken images, and unreliable performance. Filtering to Verified hosts reduces trial-and-error waste, which often negates the apparent savings of cheaper unverified machines.

Trustpilot user reviews
Compare Against RunPod and Serverless API Pricing medium success

HN discussions note that RunPod and marketplace providers offer comparable or lower GPU pricing for many configurations. Vast.ai's serverless model API pricing (per-minute) can be significantly more expensive than per-token API pricing from providers like OpenRouter for inference use cases.

Hacker News discussions
Check Actual Bandwidth Before Committing to Data-Heavy Workloads medium success

Users report advertised bandwidth (e.g., 1500Mbps) can be dramatically lower in practice (~100Mbps). For data-intensive workloads, test actual transfer speeds on a short session before committing to a longer rental to avoid cost overruns on bandwidth charges.

Trustpilot user reviews
High success Medium Low / difficult

Full negotiation guide →

Vast.ai Pricing FAQ

01 How does Vast.ai pricing work?

Vast.ai operates as a GPU marketplace where independent hosts set their own prices based on supply and demand. Prices are dynamic and vary by GPU type, host reliability, location, and market conditions. You pay per second with no minimum commitment on on-demand instances. Interruptible instances are 50%+ cheaper but can be reclaimed by hosts. A $5 minimum deposit is required to start.

02 How much does an H100 cost on Vast.ai?

H100 GPUs on Vast.ai are available from approximately $1.47/hr on the marketplace, significantly cheaper than traditional cloud providers like AWS or Lambda Labs. Prices fluctuate with supply and demand — interruptible H100 instances can be even lower. For comparison, on-demand H100s on Lambda Labs list at $4.29/hr.

03 How much does an RTX 4090 cost on Vast.ai?

RTX 4090 instances on Vast.ai start from $0.29/hr for on-demand and can be lower on interruptible pricing. Stable on-demand instances typically run $0.39–$0.59/hr. This makes it one of the cheapest places to rent RTX 4090s for inference and fine-tuning workloads.

04 Does Vast.ai charge any platform fees?

Vast.ai does not charge platform fees or markups on top of host-set prices. You pay the host's listed rate directly. There are no hidden fees, though you also pay separately for storage and bandwidth at rates set by individual hosts.

05 Is Vast.ai reliable for production workloads?

Vast.ai's on-demand instances offer guaranteed uptime and are suitable for production inference. However, as a peer-to-peer marketplace, host reliability varies. Interruptible instances are not suitable for latency-sensitive or stateful production workloads. For critical workloads, on-demand instances from highly-rated hosts are recommended.

06 What is the difference between Vast.ai's Interruptible, Reserved, and On-Demand tiers?

Vast.ai offers three rental types: Interruptible instances start at $0.10/hr and are cheapest but can be preempted at any time by higher-bid jobs; Reserved instances at $0.20/hr offer a middle ground with more stability; On-Demand instances at $0.29/hr provide uninterrupted access for time-sensitive workloads. The right choice depends on how tolerant your workload is to interruption.

07 How does Vast.ai serverless pricing compare to using an AI API provider?

Vast.ai serverless model inference costs approximately $0.30/minute ($18/hour), which community users note is significantly more expensive than token-based API pricing from providers like OpenRouter. If you only need inference rather than flexible GPU access for training or custom workloads, dedicated API providers are likely to be cheaper.

08 Is Vast.ai cheaper than other GPU cloud rental platforms?

Community comparisons suggest Vast.ai is price-competitive in the GPU marketplace segment. Users have noted it offers significantly more resources per dollar than some competitors — for example, a 28-core/32GB RAM/2x RTX 3080 Ti configuration was cited as available on Vast for roughly half the cost of an equivalent Salad instance. However, competitors like RunPod are frequently mentioned as similarly priced alternatives worth benchmarking.

09 What are the three rental tiers on Vast.ai and when should I use each?

Vast.ai offers three instance types: Interruptible (starting ~$0.10/hour) for fault-tolerant workloads that can handle preemption; On-Demand (starting ~$0.30/hour) for workloads requiring reliable uptime without interruption; and Reserved (starting ~$0.20/hour) for longer-term committed capacity. Use Interruptible for batch jobs or experiments, On-Demand for active training runs, and Reserved when you need guaranteed access over a sustained period.

10 Does Vast.ai charge for bandwidth/data transfer separately?

Yes. Vast.ai charges for internet bandwidth in addition to the hourly GPU rate. Users have reported charges of approximately $2.50 per 100GB of data transferred. This can add up quickly for workloads involving large dataset downloads, model uploads, or frequent data movement between instances.

11 Will I be charged if my instance goes offline unexpectedly?

Based on user reports, yes — billing continues even when instances become inaccessible or are shut down without warning. Multiple users report being charged for hours of downtime they could not work through, and support has been reported to decline refunds. Always monitor your instance status actively and destroy instances you're not using.

12 How does Vast.ai compare in price to RunPod or other GPU marketplaces?

HN discussions suggest Vast.ai is generally competitive or cheaper than some alternatives like Salad.com and some dedicated cloud providers, but RunPod and similar marketplace providers offer comparable pricing for many GPU configurations. For inference use cases, per-token API providers like OpenRouter may be substantially cheaper than renting a full GPU instance for serverless inference workloads.

13 Is there a free trial or free tier on Vast.ai?

No. Vast.ai does not offer a free tier or free trial. Any GPU compute workload requires a funded account with prepaid credit balance. Charges begin immediately when an instance is launched.

14 How reliable are Vast.ai instances?

Reliability varies significantly by host and instance type. Many users report successful experiences, particularly for AI training workloads, but a substantial portion of reviews cite SSH connection failures, instances going offline without warning, broken container images, and GPU performance degradation by hosts. Filtering to Verified hosts and using On-Demand rather than Interruptible instances generally improves reliability.

Is this pricing incorrect? — we'll verify and update it.