DigitalOcean

DigitalOcean serves single-GPU NVIDIA shapes. Pass provider="digitalocean" on create; the AWS default stays unchanged when you omit it.

Serving shapes

`gpu_type`	Memory	GPUs per node	Regions	Good for
`H100-80GB`	80 GB	1	tor1, nyc2	7B to 32B models in bf16 on one GPU
`H200-141GB`	141 GB	1	atl1, nyc2	70B models in fp8, or long contexts, on one GPU

These are the only single-GPU H100-class shapes on Veri: AWS sells A100/H100 as whole 8-GPU nodes. Neither DigitalOcean shape supports gpu_count above 1, and neither is a training shape (veri train on DigitalOcean uses H100-80GB x1 or x8).

Creating a deployment

client.deployments.create(
    model="Qwen/Qwen3-8B",
    source="huggingface",
    name="qwen3-8b-do",
    provider="digitalocean",
    gpu={"gpu_type": "H200-141GB", "gpu_count": 1},
    min_replicas=1,
)

veri deploy Qwen/Qwen3-8B --name qwen3-8b-do \
  --provider digitalocean --gpu-type H200-141GB --gpu-count 1

The region is chosen for you from the regions that sell the shape, so there is no region field to set.

Qwen3 models think before they answer. With a small max_tokens the visible content can come back empty because the reply is still inside reasoning_content; allow a few hundred tokens, or pass "chat_template_kwargs": {"enable_thinking": false} to skip the thinking step.

Capacity and quota

GPU Droplets come from a shared pool with no reservations, and 1x H100 capacity in particular comes and goes by the minute. When no region has the shape, the create fails before anything is billed with CAPACITY_UNAVAILABLE, and the error names the shapes and providers that do have capacity right now. A QUOTA_EXCEEDED error means the account’s GPU Droplet limit is reached; stop an idle deployment or pick another provider.

veri gpu list                 # the catalog: shapes, node counts, $/GPU-hour
veri gpu describe H200-141GB  # one shape's specs and current price

Deployments that back an agent or a monitor should set min_replicas to at least 1. A DigitalOcean replica that scales to zero wakes from a cold droplet create, and requests during the wake return 503 until the new droplet serves.

Billing

Billing starts when the droplet is created and stops when the replica is destroyed, per second, at the rate veri gpu list shows. A deployment at min_replicas=0 with no traffic costs nothing while scaled to zero. See Cost for the full model.

Get started

Training

Deployments

Evaluations

Agents

CLI

Volumes

Webhooks

Secrets

Billing

Serving shapes

Creating a deployment

Capacity and quota

Billing

Where to go next

All serving providers

AWS

​Serving shapes

​Creating a deployment

​Capacity and quota

​Billing

​Where to go next

All serving providers

AWS

Serving shapes

Creating a deployment

Capacity and quota

Billing

Where to go next