Cheapest eight-GPU node.
prices as of · re-ranked daily · 32 qualifying machines
Eight GPUs on one board is the shape a training run and a large model both want, because the GPUs talk to each other over the node's own interconnect instead of over the network. 32 configurations of eight GPUs or more qualify, from $0.173/GPU-hr a GPU-hour. The whole-machine price is beside each row, since that is the number on the invoice. Prices are the on-demand rate for one GPU for one hour, in each provider's own currency converted to dollars, so a node with eight GPUs is ranked at what one of its GPUs costs. Storage, egress and tax are extra unless the row says otherwise.
Today's pick is 8V100.48V · 8x Tesla V100 16GB at Verda (DataCrunch), $0.173/GPU-hr, read 17 Sept 2026.
- 1
Verda (DataCrunch)8V100.48V · 8x Tesla V100 16GB
8x V100 16GB · 16 GB each · 48 vCPU · 180 GB RAM · disk not stated · Verda (DataCrunch) at $0.173Dedicated Hardware Instance · storage billed separately · NVLink up to 50GB/s between GPUs
$0.173/GPU-hr
- 2
Vultrvcg-a16-48c-496g-128vram
8x A16 · 16 GB each · 48 vCPU · 496 GB RAM · 1.5 TB · Vultr at $0.471Cloud GPU · $2750/mo if kept a whole month
$0.471/GPU-hr
- 3Verda (DataCrunch)
8A6000.80V · 8x RTX A6000 48GB
8x RTX A6000 · 48 GB each · 80 vCPU · 480 GB RAM · no disk included · Verda (DataCrunch) at $0.592Dedicated Hardware Instance · storage billed separately · 50GB/s between GPUs
$0.592/GPU-hr
- 4
Lambda8x NVIDIA Tesla V100
8x V100 16GB · 16 GB each · 88 vCPU · 448 GB RAM · 5.8 TB · Lambda at $0.790Priced per GPU-hour, billed by the minute · storage included in the instance
$0.790/GPU-hr
- 5Verda (DataCrunch)
8RTX6000ADA.80V · 8x RTX 6000 Ada 48GB
8x RTX 6000 Ada · 48 GB each · 80 vCPU · 480 GB RAM · no disk included · Verda (DataCrunch) at $1.09Dedicated Hardware Instance · storage billed separately · 50GB/s between GPUs
$1.09/GPU-hr
- 6
CoreWeaveNVIDIA L40
8x L40 · 48 GB each · nvlink · 128 vCPU · 1024 GB RAM · 7.7 TB · CoreWeave at $1.25$10 an hour for the whole node · egress and VPC free · IPv4 and storage billed separately
$1.25/GPU-hr
- 7Verda (DataCrunch)
8A100.40S.176V · 8x A100 SXM4 40GB
8x A100 40GB SXM · 40 GB each · 176 vCPU · 960 GB RAM · no disk included · Verda (DataCrunch) at $1.25Dedicated Hardware Instance · storage billed separately · 600 GB/s between GPUs
$1.25/GPU-hr
- 8Verda (DataCrunch)
8L40S.160V · 8x L40S 48GB
8x L40S · 48 GB each · 160 vCPU · 480 GB RAM · no disk included · Verda (DataCrunch) at $1.44Dedicated Hardware Instance · storage billed separately · 50GB/s between GPUs
$1.44/GPU-hr
- 9Vultr
vbm-64c-2048gb-8-l40-gpu
8x L40S · 48 GB each · 64 vCPU · 2048 GB RAM · 480 GB · Vultr at $1.67Bare metal · $12000/mo if kept a whole month
$1.67/GPU-hr
- 10Verda (DataCrunch)
8A100.176V · 8x A100 SXM4 80GB
8x A100 80GB SXM · 80 GB each · 176 vCPU · 960 GB RAM · no disk included · Verda (DataCrunch) at $1.71Dedicated Hardware Instance · storage billed separately · 600 GB/s between GPUs
$1.71/GPU-hr
The number beside each machine is what one of its GPUs costs for one hour. The whole machine bills $1.38/hr at the top of this list, and more or less further down in proportion to how many GPUs it holds, so read the configuration before the rate: $0.173 a GPU-hour on an eight-GPU node is eight times that on the invoice.
Frequently asked
- What is the cheapest of the eight-GPU nodes right now?
- 8V100.48V · 8x Tesla V100 16GB at Verda (DataCrunch), $0.173/GPU-hr. The whole machine bills $1.38/hr, and it is 8x V100 16GB. Read on 17 Sept 2026.
- Which provider shows up most in this ranking?
- Verda (DataCrunch) holds 6 of the top 10 places here.
- What does the price actually cover?
- The on-demand rate for one GPU for one hour, in the provider's own currency converted to dollars. Storage, egress and tax are extra unless the row says otherwise, and reserved or committed rates are never indexed here: a price that needs a month's commitment is not a standard recurring rate, and it would beat every on-demand rate on this page.
- How is this ranked, and how current is it?
- Prices are read every morning from 10 providers' own public pages and keyless APIs, and this list recomputes with them. A machine whose provider was last read more than three days ago, or that the provider shows as unavailable, keeps its place but cannot be first. Last refresh: 17 Sept 2026.