diff --git a/vm_images/linux-amd64/linux-amd64.pkr.hcl b/vm_images/linux-amd64/linux-amd64.pkr.hcl
index fd711d1..905096c 100644
--- a/vm_images/linux-amd64/linux-amd64.pkr.hcl
+++ b/vm_images/linux-amd64/linux-amd64.pkr.hcl
@@ -7,6 +7,15 @@ packer {
   }
 }
 
+variable "iam_instance_profile" {
+  type        = string
+  description = "EC2 instance profile with permission to pull xgb-ci.gpu from ECR."
+  validation {
+    condition     = length(trimspace(var.iam_instance_profile)) > 0
+    error_message = "An instance profile with ECR pull permissions is required."
+  }
+}
+
 locals {
   ami_name_prefix = "xgboost-ci"
   image_name      = "RunsOn worker with Ubuntu 24.04 AMD64 + CUDA driver 580"
@@ -34,20 +43,21 @@ source "amazon-ebs" "runs-on-linux-amd64" {
   associate_public_ip_address = true
   communicator                = "ssh"
   instance_type               = "g4dn.xlarge"
+  iam_instance_profile        = var.iam_instance_profile
   region                      = "${local.region}"
   ssh_timeout                 = "10m"
   ssh_username                = "ubuntu"
   ssh_file_transfer_method    = "sftp"
   user_data_file              = "setup_ssh.sh"
   launch_block_device_mappings {
-    device_name = "/dev/sda1"
-    volume_size = "${local.volume_size}"
-    volume_type = "gp3"
+    device_name           = "/dev/sda1"
+    volume_size           = "${local.volume_size}"
+    volume_type           = "gp3"
     delete_on_termination = true
   }
-  aws_polling {   # Wait up to 1 hour until the AMI is ready
+  aws_polling { # Wait up to 1 hour until the AMI is ready
     delay_seconds = 15
-    max_attempts = 240
+    max_attempts  = 240
   }
   snapshot_tags = {
     Name      = "${local.image_name}"
@@ -76,4 +86,11 @@ build {
     pause_before = "1m0s"
     script       = "bootstrap.sh"
   }
+
+  provisioner "shell" {
+    script = "preload_gpu_image.sh"
+    environment_vars = [
+      "AWS_DEFAULT_REGION=${local.region}",
+    ]
+  }
 }
diff --git a/vm_images/linux-amd64/preload_gpu_image.sh b/vm_images/linux-amd64/preload_gpu_image.sh
new file mode 100644
index 0000000..3406312
--- /dev/null
+++ b/vm_images/linux-amd64/preload_gpu_image.sh
@@ -0,0 +1,23 @@
+#!/bin/bash
+# Cache Docker layers in the AMI to avoid downloading/extracting them on every job.
+# Remove this step if runners gain a persistent image cache or stop using xgb-ci.gpu.
+set -euo pipefail
+
+registry=492475357299.dkr.ecr.us-west-2.amazonaws.com
+image="${registry}/xgb-ci.gpu:main"
+
+# Use the builder's instance profile and an ephemeral Docker config so ECR tokens
+# are not baked into the AMI. CI must still pull its requested tag to pick up updates.
+docker_config=$(mktemp -d)
+trap 'sudo rm -rf "$docker_config"' EXIT
+sudo systemctl start docker
+aws ecr get-login-password --region "$AWS_DEFAULT_REGION" |
+  sudo docker --config "$docker_config" login --username AWS --password-stdin "$registry"
+sudo docker --config "$docker_config" pull "$image"
+sudo docker run --rm --pull=never --gpus all --entrypoint nvidia-smi "$image"
+# Record the digest and size in the build log to identify what the snapshot contains.
+sudo docker image inspect --format '{{json .RepoDigests}} {{.Size}}' "$image"
+
+# Preserve the downloaded layers, but stop writes before Packer snapshots the disk.
+sudo systemctl stop docker.service docker.socket containerd.service
+sync