diff --git a/inspector/inspector.py b/inspector/inspector.py
index ca65a08..717bdcb 100644
--- a/inspector/inspector.py
+++ b/inspector/inspector.py
@@ -3,6 +3,7 @@
 preserve memory for running the benchmarks.
 All other modules should be imported lazily, where needed.
 """
+from concurrent.futures import ThreadPoolExecutor
 import click
 import lib
 import logging
@@ -126,32 +127,71 @@ def start(ctx, exclude, start_only):
         runner.create(vendor, {}, RESOURCE_OPTS.get(vendor) | dict(instance=server, instance_opts=instance_opts))
 
 
-@cli.command()
-@click.pass_context
-def cleanup(ctx):
+def cleanup_task(vendor, server, data_dir):
     from datetime import datetime
     from sc_runner import runner
 
-    for srv in servers():
-        vendor = srv.vendor_id
-        server = srv.api_reference
-        data_dir = os.path.join(ctx.parent.params["repo_path"], "data", vendor, server)
-        tasks = list(lib.get_tasks(vendor))
-        if not tasks:
+    tasks = list(lib.get_tasks(vendor))
+    if not tasks:
+        return
+
+    # see if we have to destroy the resources in the Pulumi stack
+    destroy = ""
+
+    start_times = []
+    already_ended = []
+    for task in tasks:
+        meta = lib.load_task_meta(task, data_dir=data_dir)
+        if not meta.start:
             continue
-        start_times = []
-        # get the newest start time
-        for task in tasks:
-            meta = lib.load_task_meta(task, data_dir=data_dir)
-            if not meta.start:
-                continue
-            start_times.append(meta.start)
-        if start_times:
-            last_start = max(start_times)
-            # destroy the stack after a given amount of time
-            if datetime.now() - lib.DESTROY_AFTER >= last_start:
-                logging.info(f"Destroying {vendor}/{server}, last task start date: {last_start}")
-                runner.destroy(vendor, {}, RESOURCE_OPTS.get(vendor) | dict(instance=server))
+        start_times.append(meta.start)
+        if meta.end:
+            if meta.end >= meta.start:
+                already_ended.append(True)
+        else:
+            already_ended.append(False)
+
+    # if all tasks have already finished, we can destroy the stack
+    if already_ended and all(already_ended):
+        destroy = f"Destroying {vendor}/{server}, all tasks have finished"
+
+    # if DESTROY_AFTER has already passed since the newest start time, we should destroy the stack/instance
+    if start_times:
+        last_start = max(start_times)
+        if datetime.now() - lib.DESTROY_AFTER >= last_start:
+            destroy = f"Destroying {vendor}/{server}, last task start date: {last_start}"
+
+    if destroy:
+        # In order not to cause unnecessary locks in Pulumi, we first get the stack's resources to see if
+        # it's already empty, and in that case, we don't destroy it.
+        try:
+            stack = runner.get_stack(vendor, {}, RESOURCE_OPTS.get(vendor, {}) | dict(instance=server))
+        except AttributeError:
+            # this vendor is not yet supported
+            return
+        resources = stack.export_stack().deployment.get("resources", [])
+        if len(resources) <= 1:
+            # a non-existent stack will have zero, a clean (already destroyed) stack should have exactly one
+            # resource (the Pulumi Stack itself). If we can see either of these, we have nothing to clean up.
+            logging.debug(f"Pulumi stack for {vendor}/{server} has {len(resources)} resources, no cleanup needed")
+            return
+        logging.info(destroy)
+        runner.destroy(vendor, {}, RESOURCE_OPTS.get(vendor) | dict(instance=server))
+
+
+@cli.command()
+@click.pass_context
+@click.option("--threads", type=int, default=32, show_default=True,
+              help="Number of threads to run Pulumi concurrently. Each thread consumes around 60 MiB of RAM.")
+def cleanup(ctx, threads):
+    from sc_runner import runner
+    with ThreadPoolExecutor(max_workers=threads) as executor:
+        for srv in servers():
+            vendor = srv.vendor_id
+            server = srv.api_reference
+            data_dir = os.path.join(ctx.parent.params["repo_path"], "data", vendor, server)
+            # process this in a thread as getting Pulumi state is very slow
+            executor.submit(cleanup_task, vendor, server, data_dir)
 
 
 @cli.command()