A resource-already-exists error means the cloud API sees an object that the current Terraform state does not manage at the expected address. It may be a GPU instance, network, bucket, cluster, secret container, or inference endpoint. Deleting it merely to silence Terraform can create downtime or data loss.

Establish ownership first

Typical causes include manual creation, a lost or wrong state file, another stack owning the object, a configuration rename, the wrong account or workspace, or an interrupted apply.

terraform workspace show
terraform state list
terraform plan

Confirm the environment, backend, account, project, and region before changing state.

Import an unmanaged resource

Write matching configuration and use an import block:

import {
  to = aws_s3_bucket.model_artifacts
  id = "production-model-artifacts"
}

Then inspect a plan. Import establishes state ownership; it does not prove configuration matches production.

Move an existing address

If Terraform already manages the object under an old address, use a moved block or reviewed state move. Do not import one remote object into two addresses.

Delete only when ownership is clear

A collision is not evidence that the existing object is disposable. Before removal, inspect active inference traffic, DNS, model artifacts, lifecycle rules, GPU reservations, network exposure, encryption keys, and downstream queues or evaluation jobs.

Prevent drift

• Use one authoritative remote backend per environment.
• Serialize applies and retain state locking.
• Give each stack a documented ownership boundary.
• Import manual resources before further management.
• Pin module/provider versions.
• Review destructive changes in CI.
• Use documented state moves during refactors.

AI Deployment & Hosting covers environment design, AI Operations covers production safeguards, and Terraform vs Pulumi covers infrastructure workflow choices.

Safe resolution sequence

1. Stop repeated applies.
2. Confirm backend, workspace, and cloud account.
3. Identify the real owner.
4. Choose import, state move, or deliberate deletion.
5. Back up state and critical data.
6. Review replacement, cost, and exposure in the plan.
7. Apply through the approved pipeline.

The goal is a single accurate mapping between configuration, state, and real AI infrastructure—not merely a green command.