GPU Cloud Provider · Not specified
Vast.ai
Vast.ai provides high-performance computing solutions geared primarily towards AI, ML, and HPC workloads. They offer customizable, on-demand GPU clusters that enable users to scale their computing resources efficiently, securely, and cost-effectively.
GPUs
7
Founded
Not specified
Countries
20
Data Centers
1
Uptime SLA
Not published (best-effort marketplace model)
Team Size
11-50
GPU Marketplace
AMD Instinct MI100 MI100On-Demand
NVIDIA A100 80GB SXMOn-Demand
NVIDIA GeForce RTX 5090 RTX 5090On-Demand
AMD Radeon RX 7900 XT RX 7900 XTOn-Demand
AMD RX 7900 XTX XTXOn-Demand
Company Profile
Company TypeMarketplace/Aggregator
Provider TypeMarketplace
FoundedNot specified
HeadquartersNot specified
Legal EntityVast.ai, Inc.
FundingSeed
Team Size11-50
Infrastructure
GPU FleetNVIDIA H100 SXM, NVIDIA H100 PCIe, NVIDIA A100 80GB, NVIDIA A100 40GB, NVIDIA RTX 4090, NVIDIA RTX 3090, NVIDIA A40, NVIDIA A6000, NVIDIA L40, NVIDIA L40S, NVIDIA V100, NVIDIA RTX 3080, AMD Instinct MI100, NVIDIA RTX 4080, NVIDIA A10, NVIDIA A30
Total GPU CapacityNot disclosed; tens of thousands of GPUs across marketplace hosts
Network FabricConverged Network using NVIDIA Spectrum SN5600/SN5610 leaf switches, supports RDMA over NFS
ConnectivityDependent on specific deployments and configurations
StorageNVMe, Dense NVMe capacity options, NFS, S3-compatible storage
Data Center TierVaries by host; ranges from residential to Tier 3 colocation facilities
Bare MetalYes, direct access to host machines with minimal virtualization overhead in many listings
AvailabilityGA (General Availability)
StartupResearchHobbyistEnterpriseAI/ML Developers
Compute & Deployment
On-DemandYes (on-demand rentals available at fixed host-listed prices)
Spot / InterruptibleYes (interruptible instances available at lower cost; hosts can reclaim hardware)
Reserved InstancesNo
Bare MetalYes (hosts list bare metal machines; renters get direct hardware access)
VM-BasedNo (container-based, not traditional VM isolation)
Container-BasedYes (Docker)
KubernetesNo
Serverless GPUNo
Spin-Up TimeUnder 2 minutes (typically 30-90 seconds for container launch)
TerraformYes (community provider)
GPU Hardware
Latest GenH100 SXM, H100 PCIe, H200, L40S, RTX 4090, RTX 4080
Legacy SupportA100, A40, A10, V100, RTX 3090, RTX 3080, T4
Multi-GPU NodesYes (up to 8x per node)
Max GPUs/Node8
NVLinkYes (NVLink on SXM nodes)
PCIe vs SXMBoth PCIe and SXM
Pricing Model
Per HourYes (primary billing unit)
Per MinuteYes (billed per minute)
SubscriptionNo
Reserved DiscountNo
Spot DiscountYes (interruptible instances available at lower rates; discounts vary by market supply/demand)
Public PricingYes
Hidden FeesNone disclosed
Pay-as-you-goYes
Credit SystemYes (prepaid credits)
Performance & Scaling
Multi-Node TrainingLimited (manual setup required, no managed multi-node orchestration)
Elastic ScalingManual only
Auto ScalingNo
InfiniBandNot disclosed (depends on individual host hardware)
NVSwitchNot disclosed (host-dependent)
SLANot published (best-effort marketplace model)
Perf IsolationPartial (bare metal access on some hosts, but marketplace model means variable isolation)
Noisy NeighborPartial (depends on host configuration; no platform-wide guarantee)
Developer Experience
OnboardingDeploy in under 5 minutes via web UI; account creation and first instance launch is fast with credit card
FrameworksTensorFlow, PyTorch
SDK LanguagesPython
CLI ToolingFull CLI (vastai) with instance management, SSH tunneling, file copy, and search/filter capabilities
JupyterVia pre-configured Docker templates with Jupyter; SSH port forwarding also supported
TemplatesPyTorch, TensorFlow, CUDA, Jupyter Notebook, Stable Diffusion, LLM inference, ComfyUI
Model MarketplaceNone built-in; users bring their own models or use Docker Hub images
DocumentationComprehensive docs with API reference, CLI guide, and tutorials; community-supported knowledge base
API FeaturesUnspecified APIs, likely include RESTful services and management interfaces for integration with existing tools
Security & Compliance
SecurityEnterprise-grade security,Zero Trust framework,Hard multi-tenancy isolation
ComplianceEnterprise-grade security, Zero Trust framework, SLAs (Service Level Agreements)
Established marketplace since ~2018Large active user community on DiscordWidely referenced in AI/ML cost optimization discussionsTransparent public pricing and host ratings system
Data Center Locations
Coverage
CountriesUnited States, Germany, France, Netherlands, United Kingdom, Canada, Australia, Japan, Singapore, Finland, Sweden, Norway, Poland, Czech Republic, Romania, Ukraine, Russia, Brazil, South Africa, India
CitiesNot disclosed
Multi-Region FailoverNo (peer-to-peer marketplace, no managed failover)
North AmericaEuropeAsia-PacificSouth America
Compliance Regions
EU Data ResidencyYes (various EU member state hosts, specific cities not disclosed)
US Gov CloudNo
India RegionYes (host-provided, specific city not disclosed)
Datacenter Locations
Key Strengths
Lowest market prices via peer-to-peer auction model
Massive GPU variety including consumer and enterprise cards
No vendor lock-in with Docker-based deployments
Supports both on-demand and interruptible pricing
Access to rare or niche GPUs not available on hyperscalers
Known Limitations
No formal uptime SLA; hosts can reclaim machines
Variable host reliability and network quality
No managed storage or persistent volumes across instances
Security posture varies by host; not SOC 2 certified
Limited enterprise support and onboarding assistance
AMD GPU support is limited compared to NVIDIA offerings
Not suitable for compliance-sensitive workloads
Additional Information
Support Options
24/7 expert support, SLA-backed support structures
Community
Active Discord server with thousands of members; community forums and Reddit presence
Green Energy
Not disclosed; varies by individual host
Core Proposition
Peer-to-peer GPU marketplace that aggregates idle consumer and datacenter GPUs from independent hosts, offering the lowest-cost GPU compute through auction-based pricing.
Payment Methods
Credit CardCrypto
Last updated March 2026. Information subject to change.