diff --git a/litellm/proxy/middleware/in_flight_requests_middleware.py b/litellm/proxy/middleware/in_flight_requests_middleware.py
index 0172c93511d..278a585d086 100644
--- a/litellm/proxy/middleware/in_flight_requests_middleware.py
+++ b/litellm/proxy/middleware/in_flight_requests_middleware.py
@@ -83,10 +83,12 @@ class InFlightRequestsMiddleware:
         return InFlightRequestsMiddleware._gauge
 
 
-# Statuses the proxy uses when it declines to serve. A response only counts once
-# the request is also marked as shed by this proxy, since litellm forwards
-# upstream 429s with the same status.
-_SHED_STATUSES: Final = frozenset({429, 503})
+# The status this proxy uses when it declines load, and only counted once the
+# request is also marked as shed here, since litellm forwards upstream 429s with
+# the same status. The proxy's 503s are fail-closed budget rejections, a
+# dependency being unreachable rather than this pod at capacity, so counting
+# them would blur the throttle-vs-scale signal the metric exists to give.
+_SHED_STATUSES: Final = frozenset({429})
 
 
 def _record_shed_response(status: int) -> None:
diff --git a/tests/test_litellm/proxy/middleware/test_in_flight_requests_middleware.py b/tests/test_litellm/proxy/middleware/test_in_flight_requests_middleware.py
index 0ea0f53762f..9687a838c81 100644
--- a/tests/test_litellm/proxy/middleware/test_in_flight_requests_middleware.py
+++ b/tests/test_litellm/proxy/middleware/test_in_flight_requests_middleware.py
@@ -98,10 +98,11 @@ def test_non_http_scopes_not_counted():
 
 
 @pytest.mark.asyncio
-@pytest.mark.parametrize("status, expected_calls", [(429, 1), (503, 1), (200, 0), (400, 0), (500, 0)])
+@pytest.mark.parametrize("status, expected_calls", [(429, 1), (503, 0), (200, 0), (400, 0), (500, 0)])
 async def test_only_shed_responses_the_proxy_itself_produced_are_counted(status, expected_calls):
-    """A 500 is the proxy failing, not declining. Counting it would blur the
-    signal an operator uses to decide between throttling and scaling out."""
+    """A 500 is the proxy failing, not declining, and the proxy's 503s are
+    fail-closed budget rejections raised when a dependency is unreachable.
+    Counting either would blur the throttle-vs-scale signal."""
     from unittest.mock import MagicMock, patch
 
     from litellm.proxy.common_utils.request_pressure_metrics import mark_request_shed_by_proxy