How a theory about AI scaling became the world's largest GPU cloud. The founding story, intellectual origins, and mission of Vast.ai.
01 / ANALYST READ
Our read
Vast.ai aggregates GPU supply from multiple underlying operators, offering breadth across hardware types and pricing tiers. Useful for buyers who value flexibility and competitive pricing over consistency in the underlying hardware or provider.
Capacity source and provenance vary by listing. Teams with strict compliance or performance requirements should verify before committing at scale.
Confidence: MediumEvidence reviewed: 4 checks
02 / OVERVIEW
Company overview
Operator class
Marketplace
GPU models tracked
RTX3090, A100, B300 and 7 more
Services
Colocation, GPU compute, Managed services and 1 more
03 / CAPACITY
GPU lineup by use case
Tracked GPU SKUs grouped by the workload they fit best.
Training
For multi-node runs where interconnect and cluster availability decide the outcome.
A100
QuoteClaim to publish
VRAM 80 GBClass SXM
Best for Mature training stacks that value known performance and broad tooling support.
B200
QuoteClaim to publish
VRAM 192 GBClass Blackwell
Best for Frontier-scale training and teams optimizing around the newest platform.
H100
QuoteClaim to publish
VRAM 80 GBClass SXM
Best for Serious model training where time-to-result outweighs hourly cost.
H200
QuoteClaim to publish
VRAM 141 GBClass Hopper
Best for Large-scale training where memory bandwidth decides the outcome.
Inference
For predictable production serving, latency targets, and steady utilisation.
RTX3090
QuoteClaim to publish
Best for RTX3090 workloads.
B300
QuoteClaim to publish
Best for B300 workloads.
L40S
QuoteClaim to publish
VRAM 48 GBClass Ada
Best for Cost-effective production inference and image generation at scale.
RTX4090
QuoteClaim to publish
Best for RTX4090 workloads.
Fine-tuning
For smaller adaptation jobs, evaluation loops, and budget-controlled development.
RTXA6000
QuoteClaim to publish
VRAM 48 GBClass Ampere
Best for Development and fine-tuning with 48 GB of memory at workstation pricing.
04 / TRUST
Verification
Evidence that there's a real operator behind the offer.
Indicator
Marketplace classification
Vast.ai aggregates GPU capacity from multiple underlying providers through a marketplace model.
Indicator
Supply variability
Capacity source and underlying hardware vary by listing. Review individual offer terms before committing to extended runs.
05 / COMMERCIALS
Pricing
Indicative published rates. Final cluster pricing depends on commit length, storage, networking, and capacity.
Hourly rates
$0.14 – $40 /hr
Across 4 published SKUs spanning RTX3090, A100, B300, L40S.
07 / NEXT
Talk to Vast.ai
viabandwidth does not sit between you and the provider. Go straight to their site.
vast.ai
Pricing pages, quote forms, and sales routes live on the provider’s own site.
Profile pages, buyer guides, and model explainers stay open. Vendor packages cover what shows on this profile; verification stays independent and is never for sale.