diff --git a/Makefile b/Makefile
index b7fd838f..10c080e5 100644
--- a/Makefile
+++ b/Makefile
@@ -23,6 +23,7 @@ UNIT_TESTS ?= tests.test_session_lifecycle tests.test_proto_codec tests.test_for
 # Only forward BASE_URL to e2e when the user supplied it. The Makefile default
 # is for local CLI usage; e2e should start its own backend by default.
 TRAINING_TEST_BASE_URL ?= $(if $(filter environment command line,$(origin BASE_URL)),$(BASE_URL),)
+UNIT_TESTS += tests.test_dashboard_kubernetes
 TRAINING_TEST_EXTRA ?= gpu
 TRAINING_TEST_ARGS ?=
 PIGLATIN_TEST_PYTHONPATH ?= examples/sft/pig-latin
diff --git a/src/server/dashboard/__init__.py b/src/server/dashboard/__init__.py
new file mode 100644
index 00000000..e69de29b
diff --git a/src/server/dashboard/kubernetes.py b/src/server/dashboard/kubernetes.py
new file mode 100644
index 00000000..f3cfebce
--- /dev/null
+++ b/src/server/dashboard/kubernetes.py
@@ -0,0 +1,101 @@
+import asyncio
+import json
+import os
+from functools import cache
+from typing import TypedDict
+
+READ_TIMEOUT_SECONDS = 6
+
+
+class Cluster(TypedDict):
+  available: bool  # whether pods could be listed
+  pods: list[dict]
+  nodes: list[dict]
+  workloads: list[dict] | None  # None when the scheduler's Workloads cannot be listed
+  errors: list[str]
+
+
+def namespace() -> str:
+  return os.getenv("OPEN_RL_WORKER_NAMESPACE") or os.getenv("POD_NAMESPACE") or "openrl-system"
+
+
+def problem(status: dict) -> str | None:
+  if status.get("phase") == "Failed":
+    return f"Failed: {': '.join(p for p in (status.get('reason'), status.get('message')) if p) or 'see logs'}"
+  for cs in status.get("containerStatuses") or []:
+    waiting, ended = cs["state"].get("waiting"), cs["state"].get("terminated")
+    if waiting and waiting.get("reason") not in (None, "ContainerCreating", "PodInitializing"):
+      return f"{waiting['reason']}: {waiting.get('message') or ''}".strip(": ")
+    if ended and (ended.get("exitCode") or ended.get("reason") not in (None, "Completed")):
+      return f"{ended.get('reason') or 'Terminated'}: exit code {ended.get('exitCode')}" + (f" — {ended['message']}" if ended.get("message") else "")
+  if status.get("phase") == "Pending":
+    unscheduled = [c for c in status.get("conditions") or [] if c["type"] == "PodScheduled" and c["status"] != "True"]
+    return f"Unschedulable: {unscheduled[0].get('message') or 'no node available'}" if unscheduled else "Pending"
+  return None
+
+
+def gpu_capacity(node: dict, slices: list[dict]) -> int:
+  """A node using DRA publishes no nvidia.com/gpu capacity; its GPUs are the devices in its NVIDIA ResourceSlices. Only
+  each pool's newest generation counts: while the driver restarts, old and new generations coexist."""
+  name, capacity = node["metadata"]["name"], int((node["status"].get("capacity") or {}).get("nvidia.com/gpu", 0))
+  mine = [s["spec"] for s in slices if s["spec"].get("nodeName") == name and s["spec"].get("driver") == "gpu.nvidia.com"]
+  newest: dict[str, int] = {}
+  for pool in (spec["pool"] for spec in mine):
+    newest[pool["name"]] = max(newest.get(pool["name"], 0), pool["generation"])
+  devices = sum(len(spec.get("devices") or []) for spec in mine if spec["pool"]["generation"] == newest[spec["pool"]["name"]])
+  return max(capacity, devices)
+
+
+@cache
+def kubernetes_api() -> tuple:
+  from kubernetes import client, config
+
+  try:
+    config.load_incluster_config()
+  except Exception:
+    config.load_kube_config()
+  return client.CoreV1Api(), client.CustomObjectsApi()
+
+
+async def kubernetes() -> Cluster:
+  try:
+    from server.scheduler_worker_manager import worker_state
+
+    core, custom = kubernetes_api()
+  except Exception as exc:
+    return {"available": False, "pods": [], "nodes": [], "workloads": None, "errors": [f"Kubernetes unavailable: {exc}"]}
+  ns, errors = namespace(), []
+
+  async def items(name: str, call, *args) -> list[dict] | None:
+    try:
+      # Read as raw JSON, because the client's own decoding of a large pod list costs more than the request.
+      found = await asyncio.to_thread(call, *args, _preload_content=False, _request_timeout=READ_TIMEOUT_SECONDS)
+      return json.loads(found.data)["items"]
+    except Exception as exc:
+      if getattr(exc, "status", None) != 404:  # no scheduler CRD or no DRA on a local stack
+        errors.append(f"{name} read failed: {exc}")
+      return None
+
+  pods, nodes, workloads, slices = await asyncio.gather(
+    items("pods", core.list_namespaced_pod, ns),
+    items("nodes", core.list_node),
+    items("workloads", custom.list_namespaced_custom_object, "openrl.io", "v1alpha1", ns, "workloads"),
+    items("resourceslices", custom.list_cluster_custom_object, "resource.k8s.io", "v1", "resourceslices"),
+  )
+  result: Cluster = {"available": pods is not None, "pods": [], "nodes": [], "workloads": None if workloads is None else [], "errors": errors}
+  for pod in pods or []:
+    labels, statuses = pod["metadata"].get("labels") or {}, pod["status"].get("containerStatuses") or []
+    row = {"name": pod["metadata"]["name"], "node": pod["spec"].get("nodeName"), "owner": labels.get("timeslice.io/owner")}
+    result["pods"].append(row | {"problem": problem(pod["status"]), "restarts": sum(cs.get("restartCount") or 0 for cs in statuses)})
+  for node in nodes or []:
+    labels, status = node["metadata"].get("labels") or {}, node["status"]
+    row = {"name": node["metadata"]["name"], "ready": any(c["type"] == "Ready" and c["status"] == "True" for c in status.get("conditions") or [])}
+    row.update(accelerator=labels.get("cloud.google.com/gke-accelerator") or labels.get("nvidia.com/gpu.product"))
+    row["gpu_capacity"] = gpu_capacity(node, slices or [])
+    result["nodes"].append(row | {"openrl_enabled": labels.get("openrl.io/enabled") == "true"})
+  for item in workloads or []:
+    spec, status = item["spec"], item.get("status") or {}
+    deleting = bool(item["metadata"].get("deletionTimestamp"))
+    state, _ = worker_state(phase=status.get("phase") or "Pending", reason=status.get("reason") or "", node=status.get("nodeName"), deleting=deleting)
+    result["workloads"].append({"model_id": spec["modelID"], "training_kind": spec.get("trainingKind"), "state": state})
+  return result
diff --git a/tests/test_dashboard_kubernetes.py b/tests/test_dashboard_kubernetes.py
new file mode 100644
index 00000000..92c12dcd
--- /dev/null
+++ b/tests/test_dashboard_kubernetes.py
@@ -0,0 +1,40 @@
+import asyncio
+import unittest
+from unittest import mock
+
+from server.dashboard.kubernetes import gpu_capacity, kubernetes, problem
+
+
+class NodeTest(unittest.TestCase):
+  def test_a_dra_node_counts_the_gpus_in_its_resource_slices(self):
+    node = {"metadata": {"name": "n1"}, "status": {"capacity": {"cpu": "200"}}}
+
+    def slice_on(on: str, devices: int, generation: int = 2, driver: str = "gpu.nvidia.com") -> dict:
+      return {"spec": {"driver": driver, "nodeName": on, "pool": {"name": on, "generation": generation}, "devices": [{}] * devices}}
+
+    slices = [slice_on("n1", 8), slice_on("n1", 8, generation=1), slice_on("n2", 4), slice_on("n1", 4, driver="tpu.google.com")]
+    self.assertEqual((gpu_capacity(node, []), gpu_capacity(node, slices)), (0, 8))
+
+
+class ProblemTest(unittest.TestCase):
+  def test_a_pod_status_names_what_keeps_it_from_running(self):
+    crashing = {"phase": "Running", "containerStatuses": [{"state": {"waiting": {"reason": "CrashLoopBackOff", "message": "back-off 5m0s"}}}]}
+    pending = {"phase": "Pending", "conditions": [{"type": "PodScheduled", "status": "False", "message": "0/3 nodes are available"}]}
+    completed = {"phase": "Succeeded", "containerStatuses": [{"state": {"terminated": {"reason": "Completed", "exitCode": 0}}}]}
+
+    self.assertEqual(problem(crashing), "CrashLoopBackOff: back-off 5m0s")
+    self.assertEqual(problem(pending), "Unschedulable: 0/3 nodes are available")
+    self.assertIsNone(problem(completed))
+
+
+class KubernetesTest(unittest.TestCase):
+  def test_a_cluster_without_a_client_is_reported_as_unavailable(self):
+    with mock.patch("server.dashboard.kubernetes.kubernetes_api", side_effect=RuntimeError("no kubeconfig")):
+      cluster = asyncio.run(kubernetes())
+
+    self.assertFalse(cluster["available"])
+    self.assertEqual(cluster["errors"], ["Kubernetes unavailable: no kubeconfig"])
+
+
+if __name__ == "__main__":
+  unittest.main()