diff --git a/metropolis/test/e2e/suites/ha_cold/BUILD.bazel b/metropolis/test/e2e/suites/ha_cold/BUILD.bazel
new file mode 100644
index 0000000..dc738a1
--- /dev/null
+++ b/metropolis/test/e2e/suites/ha_cold/BUILD.bazel
@@ -0,0 +1,24 @@
+load("@io_bazel_rules_go//go:def.bzl", "go_test")
+
+go_test(
+    name = "ha_cold_test",
+    srcs = ["run_test.go"],
+    data = [
+        "//metropolis/node:image",
+        "//metropolis/node:swtpm_data",
+        "//metropolis/test/e2e:testimages_manifest",
+        "//third_party/edk2:firmware",
+    ],
+    tags = [
+        "resources:iops:5000",
+        "resources:cpu:3",
+        # 3x2048 for nodes plus some extra.
+        "resources:ram:7000",
+    ],
+    deps = [
+        "//metropolis/proto/common",
+        "//metropolis/test/launch",
+        "//metropolis/test/launch/cluster",
+        "//metropolis/test/util",
+    ],
+)
diff --git a/metropolis/test/e2e/suites/ha_cold/run_test.go b/metropolis/test/e2e/suites/ha_cold/run_test.go
new file mode 100644
index 0000000..6670b8f
--- /dev/null
+++ b/metropolis/test/e2e/suites/ha_cold/run_test.go
@@ -0,0 +1,86 @@
+package ha_cold
+
+import (
+	"context"
+	"fmt"
+	"testing"
+	"time"
+
+	"source.monogon.dev/metropolis/test/launch"
+	"source.monogon.dev/metropolis/test/launch/cluster"
+	"source.monogon.dev/metropolis/test/util"
+
+	cpb "source.monogon.dev/metropolis/proto/common"
+)
+
+const (
+	// Timeout for the global test context.
+	//
+	// Bazel would eventually time out the test after 900s ("large") if, for
+	// some reason, the context cancellation fails to abort it.
+	globalTestTimeout = 600 * time.Second
+
+	// Timeouts for individual end-to-end tests of different sizes.
+	smallTestTimeout = 60 * time.Second
+	largeTestTimeout = 120 * time.Second
+)
+
+// TestE2EColdStartHA exercises an HA cluster being fully shut down then
+// restarted again.
+//
+// Metropolis currently doesn't support cold startups from TPM/Secure clusters,
+// so we test a non-TPM/Insecure cluster.
+func TestE2EColdStartHA(t *testing.T) {
+	// Set a global timeout to make sure this terminates
+	ctx, cancel := context.WithTimeout(context.Background(), globalTestTimeout)
+	defer cancel()
+
+	// Launch cluster.
+	clusterOptions := cluster.ClusterOptions{
+		NumNodes:        3,
+		NodeLogsToFiles: true,
+		InitialClusterConfiguration: &cpb.ClusterConfiguration{
+			TpmMode:               cpb.ClusterConfiguration_TPM_MODE_DISABLED,
+			StorageSecurityPolicy: cpb.ClusterConfiguration_STORAGE_SECURITY_POLICY_NEEDS_INSECURE,
+		},
+	}
+	cluster, err := cluster.LaunchCluster(ctx, clusterOptions)
+	if err != nil {
+		t.Fatalf("LaunchCluster failed: %v", err)
+	}
+	defer func() {
+		err := cluster.Close()
+		if err != nil {
+			t.Fatalf("cluster Close failed: %v", err)
+		}
+	}()
+
+	launch.Log("E2E: Cluster running, starting tests...")
+
+	util.MustTestEventual(t, "Add ConsensusMember roles", ctx, smallTestTimeout, func(ctx context.Context) error {
+		// Make everything but the first node into ConsensusMember.
+		for i := 1; i < clusterOptions.NumNodes; i++ {
+			err := cluster.MakeConsensusMember(ctx, cluster.NodeIDs[i])
+			if err != nil {
+				return util.Permanent(fmt.Errorf("MakeConsensusMember(%d/%s): %w", i, cluster.NodeIDs[i], err))
+			}
+		}
+		return nil
+	})
+	util.TestEventual(t, "Heartbeat test successful", ctx, 20*time.Second, cluster.AllNodesHealthy)
+
+	// Shut every node down.
+	for i := 0; i < clusterOptions.NumNodes; i++ {
+		if err := cluster.ShutdownNode(i); err != nil {
+			t.Fatalf("Could not shutdown node %d", i)
+		}
+	}
+	// Start every node back up.
+	for i := 0; i < clusterOptions.NumNodes; i++ {
+		if err := cluster.StartNode(i); err != nil {
+			t.Fatalf("Could not shutdown node %d", i)
+		}
+	}
+	// Check if the cluster comes back up.
+	util.TestEventual(t, "Heartbeat test successful", ctx, 60*time.Second, cluster.AllNodesHealthy)
+}