On-demand GPU instances at the best price on the market. We scan live capacity and pricing across providers worldwide and route your job to the cheapest box that fits - then you provision, SSH in, and transfer your data, all from the podstack CLI.
$ podstack launch --gpu h100
TrainPods aggregates GPU capacity across providers in North America, Europe, and India and prices every launch against the whole market - so you pay a fraction of hyperscaler rates, with no egress lock-in. It's terminal-native: the podstack CLI provisions your instance, drops you into an SSH session, and transfers your data and streams your logs - without ever leaving your shell.
Every launch is routed to the lowest-priced box with stock - often ~30% under the big clouds.
Provision, SSH, and transfer data straight from the podstack CLI - no console required.
Land compute in the US, Europe, or India, near your data and users, and switch providers instantly.
One binary, from the terminal you already live in.
curl -fsSL https://podstack.ai/install | shProvision on-demand at the best live market price.
podstack launch --gpu h100Drop into your instance and sync your dataset.
podstack ssh gpu-7f3a1c
podstack cp ./data gpu-7f3a1c:/workspaceTrainPods is PodStack’s on-demand GPU product for training and fine-tuning. You provision NVIDIA GPUs, SSH in and move data entirely from the podstack CLI, and pay per second with zero egress fees.
TrainPods run up to 30% cheaper than comparable GPU clouds. Combined with per-second billing, fractional allocation and zero egress fees, the all-in cost is typically lower still. Pricing is sales-led - contact sales@podstack.ai for a quote.
TrainPods offer NVIDIA A10G 24GB, L40S 48GB, A100 40GB, A100 80GB and H100 80GB across regions in the US, EU and India, each with fractional allocation from 12.5% to 100%.
Everything runs through the podstack CLI: install it, provision a pod, then SSH in and sync datasets or weights to S3-compatible object storage at s3.podstack.ai. Because egress is free, moving data out never costs extra.
You are billed per second for exactly the GPU fraction and the seconds you use - no per-hour rounding. Allocate 12.5% of a GPU for a small job or a full H100 for large runs, and pay proportionally.
Yes. Every TrainPods GPU can be sliced from 12.5% up to 100% through PodVirt, so small experiments do not have to pay for a whole GPU while large runs can take the full card.