Serving providers

Deployments run on five providers. You pick one at create time with the provider field; it defaults to aws. There is no automatic cross-provider routing — you pick the provider explicitly, and capacity errors don’t spill to another cloud.

Providers at a glance

Provider	`gpu_type`	Memory	GPUs per node	Status
`aws` (default)	`L4-24GB`	24 GB	1	Available
`aws` (default)	`A10G-24GB`	24 GB	1	Available
`aws` (default)	`A100-80GB`	80 GB	8 (whole node)	Available
`aws` (default)	`H100-80GB`	80 GB	8 (whole node)	Available
`gcp`	`A100-80GB`	80 GB	1	Available
`gcp`	`H100-80GB`	80 GB	1	Available
`hotaisle`	`MI300X-192GB`	192 GB	1	Available
`azure`	`A100-80GB`, `H100-80GB`	80 GB	1	Coming soon
`digitalocean`	`H100-80GB`, `L40S-48GB`, `MI300X-192GB`	80/48/192 GB	1	Coming soon

Each provider sells fixed node shapes. The gpu_count you request must match one of the counts above — asking for a shape a provider doesn’t sell returns a 400 at create, before anything boots.

Rows marked coming soon are wired but feature-gated: a create returns a 400 until the gate lifts.

Choosing a provider

client.deployments.create(
    model="Qwen/Qwen2.5-0.5B-Instruct",
    source="huggingface",
    name="my-deploy",
    gpu={"gpu_type": "H100-80GB", "gpu_count": 1},
    provider="gcp",
)

Omit provider to use the default (AWS).

Checking providers from the CLI

The live version of the matrix above (including current prices) is one command:

veri gpu list

This renders the GPU catalog: one row per (GPU type, provider) with a training and a serving column — the node counts you can order for each workload (- = not offered, (soon) = wired but feature-gated off) and the per-GPU $/hr. Any filter switches to the live availability view:

veri gpu list --provider hotaisle       # live MI300X quantity
veri gpu list --gpu-type H100-80GB      # one SKU across providers
veri gpu list --provider digitalocean   # per-region droplet capacity
veri gpu list --format json             # machine-readable, incl. usd_per_hour

AVAILABLE is a real-time count (Hot Aisle live quantity, DigitalOcean per-region catalog); OFFERED (AWS) is catalog + a Spot placement-score proxy, not a launch guarantee.

Provider guides

AWS (default)

L4/A10G single GPUs and whole-node A100/H100.

GCP

Genuine single-GPU A100/H100 VMs.

Hot Aisle (AMD MI300X)

192 GB on one card — 70B-class models without tensor parallel.

Azure

Single-GPU A100/H100 — coming soon.

DigitalOcean

H100, L40S, and MI300X droplets — coming soon.

Where to go next

Hosting & GPU sizing

Pick the right GPU for your model size; cold-start expectations.

Billing

Billing states and the idle-stop warning.

Get started

Training

Deployments

Evaluations

CLI

Volumes

Providers at a glance

Choosing a provider

Checking providers from the CLI

Provider guides

AWS (default)

GCP

Hot Aisle (AMD MI300X)

Azure

DigitalOcean

Where to go next

Hosting & GPU sizing

Billing

​Providers at a glance

​Choosing a provider

​Checking providers from the CLI

​Provider guides

AWS (default)

GCP

Hot Aisle (AMD MI300X)

Azure

DigitalOcean

​Where to go next

Hosting & GPU sizing

Billing

Providers at a glance

Choosing a provider

Checking providers from the CLI

Provider guides

Where to go next