Terraform providers connect configuration to cloud APIs. If one cannot be found or installed, Terraform cannot provision the GPU compute, networking, storage, Kubernetes resources, or managed services behind an AI system.

Inspect the requirement graph

terraform {
  required_providers {
    google = {
      source  = "hashicorp/google"
      version = "~> 7.0"
    }
  }
}

The local name and provider source have different roles. Run:

terraform providers
terraform init

This reveals which root or child module introduced a requirement and whether constraints conflict.

Resolve versions deliberately

Do not delete .terraform.lock.hcl as a first response. Compare root and child constraints, the locked version, Terraform CLI version, and runner platform. If adopting a newer compatible provider is intentional, run terraform init -upgrade, review the lock-file diff, and inspect the resulting plan for changed defaults or replacements.

Registry, mirror, and checksum failures

Test from the failing runner. Check DNS, TLS interception, proxies, registry allowlists, and provider mirrors. Treat checksum failures as a supply-chain boundary; never bypass verification or install an untrusted binary to unblock deployment.

For private providers, verify hostname, namespace, type, registry authentication, and OS/CPU builds.

Installation is not cloud authentication

A provider can install successfully and still fail to access a cloud API. Check workload identity, role assumptions, region/project selection, token expiry, and required permissions separately. Prefer short-lived CI identity; never commit access keys in Terraform or variable files.

AI infrastructure safety

Provider mistakes can change GPU families and costs, select the wrong region, target the wrong Kubernetes cluster, expose inference endpoints, or replace model storage. After provider resolution:

1. run terraform validate;
2. create a plan;
3. inspect replacements, public access, regions, instance types, and state moves;
4. require approval before apply.

Connect these gates to AI Testing & Evaluation. For broader design, see AI Deployment & Hosting. For earlier failures, use Terraform init troubleshooting; for unmanaged existing objects, use Terraform resource already exists.