Contents
- Vast.AI Pricing: Overview
- Vast.ai Marketplace Model
- GPU Pricing by Model
- Vast.ai Marketplace Dynamics
- Instance Types and Pricing
- Vast.ai vs Alternatives
- Cost Optimization
- Reliability and Support
- Use Case Cost Estimates
- Real-World Cost Scenarios
- Provider Vetting Checklist
- When NOT to Use Vast.ai
- Future of Marketplace GPUs
- FAQ
- Related Resources
- Sources
Vast.AI Pricing: Overview
Vast.AI operates a decentralized GPU marketplace. Individual miners and cloud providers list spare GPU capacity, and the platform handles billing, scheduling, and support. It's more like Airbnb for GPUs than a traditional cloud provider.
Pricing is highly variable because supply comes from thousands of independent providers. The same NVIDIA H100 might rent for $1.50/hr on one provider, $2.50/hr on another, and $4.00/hr on a third, all on Vast.AI's platform simultaneously. Availability and uptime guarantees vary proportionally with price.
As of March 2026, Vast.AI claims to be the cheapest GPU rental option on the market for batch workloads that tolerate moderate reliability risk. The trade-off: no service-level agreements, limited support, and potential instance interruptions without warning.
Vast.AI Marketplace Model
How Vast.AI Works
- Independent GPU owners (miners, gaming PC owners, small datacenters) list spare capacity
- Vast.AI takes 25-30% of rental fees as platform commission
- Customers browse listings, compare price and specs, and reserve capacity
- Billing is per-second (no hourly minimums)
- Instances can be terminated by the provider with short notice
This model creates a supply glut. Providers competing on price drives rates lower than centralized clouds. But it also means inconsistent reliability: the provider might turn off their rig tomorrow if they need the GPU for something else.
Who Lists on Vast.AI
- Gaming PC owners maximizing idle capacity (cheapest, least reliable)
- Crypto miners pivoting GPU capacity when mining becomes unprofitable (variable reliability)
- Small cloud operators (Hyperstack, Crusoe) offering overflow capacity (moderate reliability)
- Data centers selling bulk capacity at discount rates (better reliability, not cheap)
Price correlates roughly with provider tier. $0.50/hr for RTX 4090 usually comes from a gamer's rig. $2.00/hr likely comes from a professional provider.
GPU Pricing by Model
Vast.AI pricing as of March 2026. Prices vary by provider, so these are representative ranges:
Consumer GPUs
| GPU | Price Range $/hr | Median $/hr | Typical Provider |
|---|---|---|---|
| RTX 3090 | $0.15 - $0.35 | $0.25 | Gaming rigs, small miners |
| RTX 4090 | $0.25 - $0.50 | $0.35 | Gaming rigs, professional small datacenter |
| RTX 6000 Ada | $0.80 - $1.50 | $1.10 | Professional builders |
Data Center GPUs
| GPU | Price Range $/hr | Median $/hr | Typical Provider |
|---|---|---|---|
| A100 (40GB) | $1.20 - $3.50 | $2.00 | Professional datacenter |
| A100 (80GB) | $1.50 - $4.00 | $2.50 | Professional datacenter |
| H100 PCIe | $1.80 - $3.00 | $2.20 | Professional datacenter, hyperscaler overflow |
| H100 SXM | $1.80 - $3.50 | $2.50 | Professional datacenter |
| H200 | $2.80-$5.00 | $3.80 | Limited availability (early 2026) |
Comparison to Fixed Cloud Providers
H100 pricing:
- Vast.AI (median): $2.20/hr
- RunPod (fixed): $1.99/hr
- Lambda (fixed): $3.78/hr (SXM)
- AWS (fixed): $6.88/hr
Vast.AI's median is slightly above RunPod. But Vast.AI's range is $1.80 to $3.50, so it's possible to find cheaper than RunPod ($1.80) depending on which provider teams pick.
Vast.AI Marketplace Dynamics
Price Discovery
When teams open Vast.AI's browser, teams see dozens of H100 listings sorted by price. Each shows:
- Provider name (or anonymous)
- GPU count and type
- VRAM
- Storage
- $/hr (per GPU)
- Uptime percentage (e.g., "99.2% uptime over 30 days")
- User reviews (1-5 stars)
- Availability (how many are in stock right now)
Cheapest listings often have lower uptime ratings or are from new/unrated providers. Most reliable listings are more expensive. This creates a risk-reward tradeoff that fixed-price clouds don't have.
Spot Pricing (Bid System)
Vast.AI allows bid-based spot rentals. Teams offer a maximum hourly rate and Vast.AI fulfills the request if any provider accepts. This can be 30-50% cheaper than on-demand on popular GPUs, but instances may terminate without warning if a provider accepts a better offer elsewhere.
Example: RTX 4090 spot bid at $0.28/hr might take 5 minutes to fulfill on a gamer's rig. But if the provider decides to use the GPU for another purpose, the instance terminates.
Instance Types and Pricing
On-Demand (Reserved by Hour)
Select a provider and reserve for 1 hour, 1 day, 1 month, or longer. Price is per-second, so charging is granular.
Example: RTX 4090 at $1.10/hr:
- 1 hour: $1.10
- 8 hours: $8.80
- 24 hours: $26.40
- 30 days (720 hrs): $792
No monthly discount (unlike AWS or Lambda). Flat per-hour rate.
Interruptible (Spot Bid)
Bid a maximum price. Vast.AI matches with the cheapest provider. Instance terminates if:
- Provider disconnects (their rig goes offline)
- A user bids higher for the same instance
- Provider reclaims the GPU for personal use
Typical discount: 30-50% off on-demand rates.
RTX 4090 spot bid at $0.25/hr (vs $0.35 on-demand). Same GPU, ~30% cheaper, with termination risk.
Bulk/Committed Pricing
Renting for 30+ days on the same provider sometimes negotiates a discount (5-15% off), but this is informal and provider-dependent. No formal contract.
Vast.AI vs Alternatives
Vast.AI vs RunPod
| Dimension | Vast.AI | RunPod |
|---|---|---|
| H100 PCIe $/hr | $1.80-$5.00 (median $2.80) | $1.99 (fixed) |
| Uptime SLA | None (varies by provider) | 99% uptime SLA |
| Support | Community forums (slow) | Email support (24-48 hrs) |
| Predictability | Low (prices fluctuate) | High (fixed rates) |
| Spot pricing | Yes (bid-based) | Yes ($0.99-$1.50/hr H100) |
| Best for | Cost-sensitive, fault-tolerant | Production workloads, stability |
RunPod is cheaper and more reliable for production. Vast.AI is cheaper for teams that find a good provider, but comes with risk.
Vast.AI vs Lambda
| Dimension | Vast.AI | Lambda |
|---|---|---|
| H100 PCIe $/hr | $1.80-$5.00 (median $2.80) | $2.86 (fixed) |
| Uptime SLA | None | 99.5% uptime SLA |
| Support | Community | Email + phone support |
| API stability | Lower (provider-dependent) | Higher (managed) |
| Best for | Batch jobs, experimenting | Production inference, training |
Lambda is more expensive but more professional. Vast.AI is cheaper if teams are willing to manage provider risk.
Vast.AI vs AWS/GCP
| Dimension | Vast.AI | AWS | GCP |
|---|---|---|---|
| H100 SXM $/hr | $1.80-$3.50 (median $2.50) | $6.88 | $11.68 |
| Uptime SLA | None | 99.9% | 99.9% |
| Support | Community | AWS support plans | GCP support plans |
| Compliance | None | SOC 2, HIPAA, FedRAMP | SOC 2, HIPAA |
| Best for | Budget-conscious | Large-scale, regulated | Large-scale, regulated |
Vast.AI is 2-3x cheaper than hyperscalers, but that's the market: boutique providers are always cheaper than hyperscalers.
Cost Optimization
Provider Selection
Browse provider reviews before booking. 98%+ uptime providers are generally safe. 95% or lower is risky. Cross-reference with user comments: "Provider was stable for 30 days straight" vs "Disconnected twice in a week."
Spot Bidding
For fault-tolerant workloads (training with checkpoints, batch processing), bid 20-30% below on-demand rates. The trade-off is worth it if the job checkpoints regularly.
Batch Scheduling
Use Vast.AI for batch jobs (training runs, preprocessing, backfilling). Use fixed-price clouds (RunPod, Lambda) for production inference where uptime SLAs matter. This hybrid approach minimizes cost while managing risk.
Machine Type Right-Sizing
Vast.AI lists RTX 4090 (24GB), A100 (40GB or 80GB), and H100 variants side-by-side. Sometimes an A100 at $2.00/hr is better than an H100 at $2.80/hr for the workload. Compare per-TFLOP and VRAM cost, not just hourly rate.
Reliability and Support
What Teams get
Vast.AI provides the platform (matching, billing, basic support). The provider supplies the hardware, internet, and uptime.
Teams inherit the provider's reliability:
- Gaming PC owner: ~95% uptime (might turn off rig at any time)
- Professional datacenter: 99%+ uptime (production-grade infrastructure)
Vast.AI's platform is stable, but instance termination is beyond Vast.AI's control.
Support
Vast.AI has community forums (no SLA) and email support (response time: 1-3 days). Not 24/7. Not production-grade.
RunPod, Lambda, and AWS all have faster support. Vast.AI is self-service + community.
When Vast.AI is Risky
- Production inference with uptime requirements
- Real-time applications (live user traffic)
- Multi-week long-running jobs (high likelihood of provider disconnect)
- Regulated workloads (compliance, healthcare, finance)
When Vast.AI is Safe
- Training with checkpointing (restart jobs cleanly)
- Batch preprocessing (restartable)
- Research and experimentation (failures are OK)
- Short jobs (< 8 hours) with multiple restarts planned
Use Case Cost Estimates
Fine-Tuning 7B Mistral (Single GPU, 10 hours)
Vast.AI median A100 (40GB) at $2.00/hr:
- 10 hours × $2.00 = $20
RunPod same A100 at $1.19/hr:
- 10 hours × $1.19 = $11.90
Vast.AI is ~67% more expensive than RunPod in this case. The provider tested was mid-tier. Cheaper providers might close that gap.
Training (8x H100 SXM, 7 days with checkpointing)
Vast.AI median H100 SXM at $2.50/hr per GPU:
- 8 GPUs × $2.50 × 168 hours = $3,360
CoreWeave 8x H100 at $49.24/hr (all-in):
- 168 hours × $49.24 = $8,272
Vast.AI wins by 59%. However, this assumes:
- No instance terminations (risky assumption)
- Provider uptime holds for a week (possible but not guaranteed)
- Spot bidding at median rates (actual prices vary)
If an instance terminates mid-job, restarting costs additional GPU hours.
Inference Serving (2M tokens/day, 4 hours GPU/day)
Vast.AI RTX 4090 spot bid at $0.35/hr:
- 4 hrs/day × $0.35 × 30 days = $42/month
RunPod RTX 4090 at $0.34/hr:
- 4 hrs/day × $0.34 × 30 days = $41/month
RunPod is cheaper and more reliable. Spot bidding on Vast.AI is risky for production inference.
Real-World Cost Scenarios
Startup Training Small Models (Weekly)
A startup training 3B-7B models, 5 hours per week for research:
Vast.AI approach:
- RTX 3090 spot bid at $0.22/hr (median)
- 5 hrs/week × $0.22 × 4 weeks = $4.40/month
- Risk: instance interruptions mid-training (need checkpointing)
RunPod approach:
- RTX 3090 on-demand at $0.22/hr
- 5 hrs/week × $0.22 × 4 weeks = $4.40/month
- No risk, stable, but more expensive
Costs are comparable between providers in this scenario; choosing RunPod offers simpler, more reliable operation.
ML Researcher Experimenting Daily
Daily experiments on H100, 2 hours per day:
Vast.AI (median H100 at $2.20/hr):
- 2 hrs/day × $2.20 × 30 days = $132/month
Lambda (H100 SXM at $3.78/hr):
- 2 hrs/day × $3.78 × 30 days = $226.80/month
Vast.AI is about 12% cheaper, but Lambda offers SLAs. A modest difference; choose based on reliability needs.
Production Inference System (24/7)
Serving a chatbot, 24/7 on RTX 4090:
Vast.AI on-demand (median $0.35/hr):
- 730 hrs/month × $0.35 = $255/month
- Risk: provider could disconnect anytime
RunPod spot ($0.34/hr historical):
- 730 hrs/month × $0.34 = $248/month
- Risk: interruption every 2-4 hours typical
- Requires fault tolerance and job restart logic
RunPod on-demand ($0.34/hr):
- 730 hrs/month × $0.34 = $248/month
- No interruption risk
Vast.AI and RunPod are comparably priced for this use case, but RunPod has the reliability advantage. Clear win for RunPod.
Provider Vetting Checklist
When evaluating a Vast.AI provider before booking:
- Uptime percentage: Look for 98%+ (green indicator). Below 95% is risky.
- Review count: More is better. 100+ reviews = established. <5 reviews = unproven.
- Star rating: 4.5+ is safe. Below 4.0 suggests problems.
- Recent reviews: Read the last 10. Look for keywords like "stable," "responsive," "quick support."
- Hardware notes: Some providers mention "gaming rig" or "mining PC." More casual deployments = higher risk.
- Pricing: Outliers (super cheap or expensive) might indicate something (unestablished provider or professional datacenter).
When NOT to Use Vast.AI
Avoid Vast.AI for:
- Production inference with uptime SLAs
- Multi-day continuous training (risk of interruption)
- Cost-critical applications where $X vs $X+10% matters less than certainty
- Regulated workloads (healthcare, finance)
- Real-time applications (stock trading, live customer systems)
Vast.AI works for:
- Research and prototyping
- Training with checkpoints
- Batch processing jobs that can restart
- Experimenting with new models
- Teams with tight budgets and fault-tolerant architectures
Future of Marketplace GPUs
Vast.AI faces headwinds. As fixed-price cloud providers (RunPod, Lambda, CoreWeave) mature and drop prices through 2026, the price advantage of marketplace GPUs narrows. Meanwhile, the reliability disadvantage remains.
By Q4 2026, prediction: Vast.AI's price advantage will be ~10-15% (instead of today's 20-30%), making the reliability tradeoff harder to justify.
Vast.AI's moat is network effects and massive provider catalog. If that erodes, boutique clouds with SLAs will win market share.
FAQ
Is Vast.AI really the cheapest GPU cloud? On paper, yes (median prices are competitive with RunPod and cheaper than Lambda). In practice, variability is high. Some providers are cheaper than RunPod; others are as expensive as AWS. Review ratings before booking.
Can I run production workloads on Vast.AI? Not recommended. No SLAs. No guarantees. Instances can terminate without warning. Use Vast.AI for batch and research. Use RunPod or Lambda for production.
What happens if my instance terminates? Work in progress is lost unless you checkpoint regularly. Vast.AI credits the unused balance; you're not charged for downtime. But you lose compute time.
How do I find reliable providers on Vast.AI? Sort by uptime percentage (98%+), sort by reviews (4+ stars), and read comments. Avoid new providers and gaming rigs if you need stability.
Can I rent Vast.AI GPUs long-term? Yes. Book a week or a month at a time. No discount for longer terms. Some providers will negotiate informal discounts for multi-week bookings if you contact them directly.
Does Vast.AI have an API? Yes. REST API for launching instances, managing billing, and querying provider lists. But it's less documented than RunPod or Lambda.
What's the minimum rental duration? Vast.AI bills per-second, so technically you can rent for 60 seconds. In practice, small rentals are impractical because setup time (SSH, loading data) eats into a short rental. Minimum practical duration is 30 minutes.
How do I know if a provider is trustworthy? Look at:
- Uptime percentage (98%+ is good, 99%+ is excellent)
- Review count (more reviews = more track record)
- Star rating (4.5+ is good)
- Comments mentioning stability
Avoid providers with <95% uptime or negative reviews about disconnections.
Related Resources
- GPU Cloud Provider Comparison
- Lambda Cloud GPU Pricing
- JarvisLabs GPU Pricing
- FluidStack GPU Pricing
Sources
- Vast.ai Official Website
- Vast.ai GPU Marketplace
- Vast.ai API Documentation
- Vast.ai Community Forum
- DeployBase GPU Pricing Tracker (provider rates observed March 21, 2026)