Docs · how to

Connect a hyperscaler

Let denv launch GPU hosts in your AWS, Google Cloud, Azure or Oracle Cloud account, with a scoped role you can revoke.

denv launches hosts in your own account. The provider bills you for the instances, as it does today, so reservations, savings plans, committed spend and credits still apply.

Every host boots from a machine image with denvd and a supported GPU driver preinstalled. denv publishes images per region and driver version; you do not build or patch them.

What denv can do in your account

Allowed Why
Launch instances tagged denv:managed=true Create hosts for your fleets
Stop and terminate instances tagged denv:managed=true Scale down, replace failed hosts
Describe instances, images, subnets, security groups, placement groups Placement and health
Read service quotas Tell you before a fleet asks for GPUs you cannot get

It cannot touch untagged instances, read your buckets, or change identity settings. Remove the role and denv loses access at once; running hosts keep serving until you remove them. See Security.

AWS

denv provider setup aws --regions us-east-1,us-west-2

This prints a CloudFormation link and your organization's external id. Create the stack in the account you want to use, copy its RoleArn output, then:

denv provider connect aws --name aws-prod \
  --role-arn arn:aws:iam::123456789012:role/denv-provider \
  --regions us-east-1,us-west-2

To use Capacity Blocks or capacity reservations, create the fleet with --mode reserved in the reservation's region and zone. denv launches into the reservation first.

Google Cloud

denv uses workload identity federation; no service-account keys leave your project.

denv provider setup gcp --project acme-inference --regions us-central1

Run the printed gcloud script, then:

denv provider connect gcp --name gcp-prod \
  --project acme-inference \
  --service-account denv-provider@acme-inference.iam.gserviceaccount.com \
  --regions us-central1

Azure

denv provider setup azure --subscription <subscription-id> --regions eastus2

The printed az script creates the resource group denv-managed, an app registration and a role assignment scoped to that group. Then:

denv provider connect azure --name azure-prod \
  --subscription <subscription-id> --tenant <tenant-id> --client-id <client-id> \
  --regions eastus2

Oracle Cloud

denv provider setup oci --tenancy <tenancy-ocid> --regions us-ashburn-1
denv provider connect oci --name oci-prod --tenancy <tenancy-ocid> --regions us-ashburn-1

Apply the policy statements setup prints in a compartment for denv before you connect.

Check GPU quota

New accounts often have no GPU quota. Check before you create a fleet:

denv offers --hardware h200 --count 8 --provider aws-prod

The output lists each shape with its hardware id, domain size, capacity mode and your remaining quota per region. A quota of 0 means you need to request an increase in the provider's console first. Hardware ids are listed in Hardware.

Create a fleet

denv fleet create h200-use1 --provider aws-prod --region us-east-1 \
  --shape p5en.48xlarge --mode reserved --min 2 --max 8

denv launches hosts up to --min. Each reports its GPUs, driver and scale-up domain when it joins:

denv host list --fleet h200-use1

Check the connection

denv provider list

A provider is ready once denv has assumed the role and read quotas in each region. Nothing is launched until a fleet asks for capacity.

Next