diff --git a/litellm/proxy/spend_tracking/daily_global_spend_rollup.py b/litellm/proxy/spend_tracking/daily_global_spend_rollup.py
index c135c7d1d9c..d50feb6f16b 100644
--- a/litellm/proxy/spend_tracking/daily_global_spend_rollup.py
+++ b/litellm/proxy/spend_tracking/daily_global_spend_rollup.py
@@ -161,6 +161,24 @@ async def _record_marker(prisma_client: "PrismaClient", marker: ReconciledThroug
     await invalidate_config_param(DAILY_GLOBAL_SPEND_RECONCILED_THROUGH_PARAM)
 
 
+async def _stored_marker(prisma_client: "PrismaClient") -> ReconciledThrough | None:
+    """The marker as another pod may have just written it, bypassing this pod's config cache."""
+    param: Final = await ConfigRepository(prisma_client).get_param(DAILY_GLOBAL_SPEND_RECONCILED_THROUGH_PARAM)
+    return None if param is None else _marker_from_param_value(param.param_value)
+
+
+async def _record_advanced(prisma_client: "PrismaClient", days: tuple[str, ...], *, scanned_at: str | None) -> None:
+    """Advance the stored marker by ``days``. Two runs can overlap (Redis unreachable, lock expired
+    on a long backfill), so the base is what is stored now, not the snapshot this run scanned from:
+    a slower run may then only add to the faster run's marker, never rewind it. Without a new scan
+    time the stored one is kept."""
+    stored: Final = await _stored_marker(prisma_client)
+    kept_scanned_at: Final = None if stored is None else stored.scanned_at
+    await _record_marker(
+        prisma_client, _advanced(stored, days, scanned_at=scanned_at if scanned_at is not None else kept_scanned_at)
+    )
+
+
 async def _db_now(prisma_client: "PrismaClient") -> _NowRow:
     rows: Final = await prisma_client.db.query_raw(_DB_NOW_SQL)
     return _NowRow.model_validate(rows[0])
@@ -203,7 +221,7 @@ async def run_daily_global_spend_reconcile(prisma_client: "PrismaClient") -> Rec
         marker: Final = await reconciled_through(prisma_client)
         return ReconcileResult(days_reconciled=done, reconciled_through=marker, failed_day=scan.days[len(done)])
     if scan.marker is not None or done:
-        await _record_marker(prisma_client, _advanced(scan.marker, done, scanned_at=scan.scanned_at))
+        await _record_advanced(prisma_client, done, scanned_at=scan.scanned_at)
     return ReconcileResult(days_reconciled=done, reconciled_through=await reconciled_through(prisma_client))
 
 
@@ -215,21 +233,16 @@ def _advanced(marker: ReconciledThrough | None, days: tuple[str, ...], *, scanne
 
 async def _reconcile_until_failure(prisma_client: "PrismaClient", scan: _PendingScan) -> tuple[str, ...]:
     for index, day in enumerate(scan.days):
-        if not await _reconcile_and_record(prisma_client, scan.marker, scan.days[: index + 1]):
+        if not await _reconcile_and_record(prisma_client, scan.days[: index + 1]):
             return scan.days[:index]
     return scan.days
 
 
-async def _reconcile_and_record(
-    prisma_client: "PrismaClient", marker: ReconciledThrough | None, done_with_this: tuple[str, ...]
-) -> bool:
+async def _reconcile_and_record(prisma_client: "PrismaClient", done_with_this: tuple[str, ...]) -> bool:
     day: Final = done_with_this[-1]
     try:
         await reconcile_day(prisma_client, day)
-        await _record_marker(
-            prisma_client,
-            _advanced(marker, done_with_this, scanned_at=None if marker is None else marker.scanned_at),
-        )
+        await _record_advanced(prisma_client, done_with_this, scanned_at=None)
     except Exception as exc:  # noqa: BLE001  # one bad day must not lose the days already done
         verbose_proxy_logger.exception("Daily global spend reconcile: day %s failed: %s", day, exc)
         return False
diff --git a/tests/test_litellm/proxy/spend_tracking/test_daily_global_spend_rollup.py b/tests/test_litellm/proxy/spend_tracking/test_daily_global_spend_rollup.py
index 9655953134a..69f06fad081 100644
--- a/tests/test_litellm/proxy/spend_tracking/test_daily_global_spend_rollup.py
+++ b/tests/test_litellm/proxy/spend_tracking/test_daily_global_spend_rollup.py
@@ -40,6 +40,10 @@ class _FakeConfigTable:
         self.rows[where["param_name"]] = data["update"]["param_value"]
         return _FakeConfigRow(where["param_name"], data["update"]["param_value"])
 
+    async def find_unique(self, *, where: dict[str, str]) -> _FakeConfigRow | None:
+        stored = self.rows.get(where["param_name"])
+        return None if stored is None else _FakeConfigRow(where["param_name"], stored)
+
 
 class _FakeDb:
     """Per-key rows are ``{date: updated_at}`` with a fake database clock that ticks per query,
@@ -68,11 +72,15 @@ class _FakeDb:
         if day in self._prisma.failing_days:
             raise RuntimeError(f"day {day} exploded")
         self._prisma.reconciled.append(day)
+        landing = self._prisma.marker_landing_on_day.get(day)
+        if landing is not None:
+            self.litellm_config.rows[DAILY_GLOBAL_SPEND_RECONCILED_THROUGH_PARAM] = landing
         return 1
 
 
 class _FakePrisma:
-    """Enough of PrismaClient for the reconcile: per-key dates, a config table, and execute_raw."""
+    """Enough of PrismaClient for the reconcile: per-key dates, a config table, and execute_raw.
+    ``marker_landing_on_day`` stores another pod's marker the moment this run rewrites that day."""
 
     def __init__(
         self, user_days: tuple[str, ...], failing_days: frozenset[str] = frozenset(), today: date = TODAY
@@ -81,6 +89,7 @@ class _FakePrisma:
         self.today = today
         self.user_rows: dict[str, str] = {d: "clock-0000" for d in user_days}
         self.failing_days = failing_days
+        self.marker_landing_on_day: dict[str, str] = {}
         self.reconciled: list[str] = []
         self.db = _FakeDb(self)
 
@@ -221,6 +230,25 @@ async def test_the_next_run_resumes_from_the_failed_day():
     assert await reconciled_through(prisma) == "2026-09-03"
 
 
+@pytest.mark.asyncio
+async def test_a_slower_overlapping_run_never_rewinds_the_marker_a_faster_run_stored():
+    """Two pods can reconcile at once (Redis unreachable, or the lock expired on a long backfill).
+    When the faster one has already stored a later marker, the slower one may only add to it. Putting
+    its own older prefix back, or dropping the scan time, would send usage reads for every day in
+    between back to the per-key table until the next run."""
+    prisma = _FakePrisma(user_days=("2026-09-01", "2026-09-02", "2026-09-03"), failing_days=frozenset({"2026-09-03"}))
+    prisma.marker_landing_on_day = {
+        "2026-09-02": '{"reconciled_through": "2026-09-14", "scanned_at": "clock-0009"}',
+    }
+
+    result = await run_daily_global_spend_reconcile(prisma)
+
+    assert result.days_reconciled == ("2026-09-01", "2026-09-02")
+    assert result.reconciled_through == "2026-09-14"
+    marker = await read_marker(prisma)
+    assert marker is not None and (marker.reconciled_through, marker.scanned_at) == ("2026-09-14", "clock-0009")
+
+
 @pytest.mark.asyncio
 async def test_a_failure_with_nothing_done_reports_the_previous_marker_and_alerts():
     """When the rewrite of a late day fails the marker must stay put and the operator must hear about it."""