The verdict
Vast.ai takes it, 5–2 on rounds.
On the published figures and the independent review sites, Vast.ai takes 5 of 8 rounds. Treat that as a starting point and check the current figures on both pages, because prices and plans move.
Choose Vast.ai if…
- • You want Vast.ai's edge on the cheapest h100 and the cheapest a100
Choose AWS if…
- • You want AWS's edge on what it costs to get your data out and how much you save by committing
Round by round
- Round 1The cheapest H100Vast.ai winsVast.ai8/10
$1.48
H100 SXM
- the lowest H100 offer on the marketplace (offers seen ran from about $2 to $19)
Source: vast.ai ↗
AWS5/10$6.88
p5.48xlarge $55.04/hr for 8x H100 = $6.88 per GPU-hour, on-demand, US East
- $5.19, against $0.90
Source: aws.amazon.com ↗
Vast.ai wins this round on what the company itself publishes.
- Round 2The cheapest A100Vast.ai winsVast.ai8/10
$0.27/hr
A100 PCIe lowest offer (grid card: from $0.27/hr, median $0.71/hr)
- the lowest A100 PCIe offer on the marketplace
Source: vast.ai ↗
AWS5/10$2.74
worked out from $21.958 an hour for a p4d.24xlarge with eight A100s
- $2.74, against $0.35
Source: instances.vantage.sh ↗
Vast.ai wins this round on independent review-site data, not on its own claims.
- Round 3What it costs to get your data outAWS winsVast.ai5/10
Bandwidth charged both ways
billed per byte sent or received, at rates that vary by host
- Bandwidth charged both ways, against First 100 GB free, then $0.09 per GB
Source: docs.vast.ai ↗
AWS8/10First 100 GB free, then $0.09 per GB
data transfer out to the internet
Source: aws.amazon.com ↗
AWS wins this round on what the company itself publishes.
- Round 4How many places it runsVast.ai winsVast.ai8/10
40+ data centers
Marketplace coverage wording: 'Prices set by supply and demand across 40+ data centers.' Homepage also says 'On-demand instances across 40+ data centers and 20,000+ GPUs.'
- host locations on a marketplace, not data-center regions
Source: vast.ai ↗
Vast.ai wins this round on what the company itself publishes.
- Round 5What customers say on TrustpilotVast.ai winsAWS5/10
1.3 out of 5
TrustScore from 429 reviews on Trustpilot
- 1.3 out of 5, against 4.1 out of 5
Source: trustpilot.com ↗
Vast.ai wins this round on independent review-site data, not on its own claims.
- Round 6How precisely you are billedTieAWS7/10
Per-second, 60-second minimum
on-demand instances are billed by the hour or second
Source: aws.amazon.com ↗
Nothing to separate them here.
- Round 7Whether GPUs can run serverlessVast.ai winsAWS5/10
No GPUs in Serverless Inference
EC2 on-demand, US East (N. Virginia), Linux
- No GPUs in Serverless Inference, against Yes
Source: docs.aws.amazon.com ↗
Vast.ai wins this round on what the published sources say.
- Round 8How much you save by committingAWS winsVast.ai5/10
Up to 50% off
marketplace GPU cloud (on-demand / interruptible / reserved)
- Up to 50% off, against Up to 72% off
Source: vast.ai ↗
AWS wins this round on what the published sources say.
Side by side
- the lowest H100 offer on the marketplace (offers seen ran from about $2 to $19)
- the lowest A100 PCIe offer on the marketplace
- the lowest rate on the marketplace, for consumer GPUs
- host locations on a marketplace, not data-center regions
- AWS Regions
- worked out from $41.528 an hour for a p5.48xlarge with eight H100s, US regions
- worked out from $47.76 an hour for a p5e.48xlarge with eight H200s
- Bandwidth charged both ways, against No egress fees
- Bandwidth charged both ways, against Free
- Up to 50% off, against Up to 60% off
- $0.13, against Free GPU notebooks
- $5.19, against $1.99
- $2.74, against $1.19
- $5.97, against $3.59
- First 100 GB free, then $0.09 per GB, against No egress fees