provider field. Each provider offers different GPU types at different price points: AWS is the default, Vast.ai is the NVIDIA marketplace spillover, DigitalOcean offers 1x and 8x H100 droplets with per-second billing, and GCP offers a single A100 80GB or H100 80GB on Flex-start VMs.
Providers at a glance
GCP training runs on Flex-start VMs: a GCP job can wait in
queued for up to about 15 minutes while Veri looks for capacity, and Veri runs one GCP GPU at a time for now. If GCP ends the VM early the job fails with provider_reclaimed (credited back) or, at the 24-hour limit, max_run_duration_reached; resume it from a checkpoint. See GCP.
All providers except AWS are gated behind feature flags. If a provider is disabled, submit returns a 400 saying the provider isn’t available yet. Run
veri gpu list to see which provider and GPU combinations are offered for training.Choosing a provider
Passprovider when creating a training job:
veri train my_job.py --provider digitalocean does the same for a script with a @training_job decorator, which sets the GPU itself.)
Or in a TOML config:
provider to use the default (AWS).
Checking live availability
Capacity is per-provider and per-region. Check what’s available right now before submitting:veri gpu list is the static catalog: which node sizes each provider sells for training and serving, and the price per GPU-hour. Add any filter (--provider, --gpu-type, --region) for the live view: one row per provider, GPU type and region with a status (AVAILABLE for a DigitalOcean region or Vast offer, OFFERED for an AWS catalog entry, UNAVAILABLE when there is no capacity) and a signal column (qty N for a live DigitalOcean or Vast count, spot N/10 for the AWS Spot placement score). AWS OFFERED is not a launch guarantee.
Pricing
Each provider has its own rate card. Check current pricing with:provisioning, before the first training step) until the job terminates and the instance is reaped. This applies to failed jobs too: GPU time actually used is billed. A job that never secures a GPU (e.g. CAPACITY_UNAVAILABLE) rents nothing and is not billed. The checkpoints and the final model a job keeps are billed separately, per GiB-month kept; see Checkpoint storage and Model storage.
Provider guides
AWS (default)
L4, A10G, and whole-node A100/H100 — the reliable NVIDIA default.
Vast.ai
Marketplace A100/H100 spillover at fluctuating prices.
DigitalOcean
Fast-provisioning 1x/8x H100 droplets with per-second billing.
GCP
A single A100 or H100 on Flex-start VMs, one GCP GPU at a time.
Where to go next
Managed training
How the training loop works across all providers.
GRPO algorithm
The default RL method — works on all providers.