diff --git a/packages/orchestrator/pkg/server/post_crash_flush_test.go b/packages/orchestrator/pkg/server/post_crash_flush_test.go
new file mode 100644
index 0000000000..0da4191658
--- /dev/null
+++ b/packages/orchestrator/pkg/server/post_crash_flush_test.go
@@ -0,0 +1,41 @@
+package server
+
+import (
+	"errors"
+	"fmt"
+	"syscall"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+)
+
+// wrapLikeCleanup reproduces the real wrapping depth a flush error reaches the
+// cleanup call site through: path flush -> NBDProvider.Close -> cleanup.Run's
+// errors.Join -> Sandbox.Close.
+func wrapLikeCleanup(inner error) error {
+	flush := fmt.Errorf("sync NBD device: %w", inner)
+	cow := fmt.Errorf("error flushing cow device: %w", flush)
+	joined := errors.Join(cow)
+	return fmt.Errorf("failed to cleanup sandbox: %w", joined)
+}
+
+func TestIsExpectedPostCrashFlushError(t *testing.T) {
+	t.Parallel()
+
+	for name, tc := range map[string]struct {
+		err  error
+		want bool
+	}{
+		"EIO deep-wrapped":   {wrapLikeCleanup(syscall.EIO), true},
+		"ENXIO deep-wrapped": {wrapLikeCleanup(syscall.ENXIO), true},
+		"EIO bare":           {syscall.EIO, true},
+		"disk full (ENOSPC)": {wrapLikeCleanup(syscall.ENOSPC), false},
+		"generic error":      {wrapLikeCleanup(errors.New("boom")), false},
+		"nil":                {nil, false},
+	} {
+		t.Run(name, func(t *testing.T) {
+			t.Parallel()
+			assert.Equal(t, tc.want, isExpectedPostCrashFlushError(tc.err))
+		})
+	}
+}
diff --git a/packages/orchestrator/pkg/server/sandboxes.go b/packages/orchestrator/pkg/server/sandboxes.go
index 6abe238390..94c30e4b16 100644
--- a/packages/orchestrator/pkg/server/sandboxes.go
+++ b/packages/orchestrator/pkg/server/sandboxes.go
@@ -2426,6 +2426,18 @@ func (s *Server) sandboxAlreadyRunning(ctx context.Context, sandboxID, execution
 	return status.Errorf(codes.AlreadyExists, "sandbox '%s' is already running on this node: %s", sandboxID, cause)
 }
 
+// isExpectedPostCrashFlushError reports whether a sandbox Close error is the
+// expected consequence of flushing an NBD device whose VM has already exited.
+// After a crash or OOM-kill the device is in error state, so the rootfs flush
+// (fsync + BLKFLSBUF on /dev/nbdX) returns EIO or ENXIO; the data is already
+// gone with the VM, so at the post-exit cleanup these are noise, not failures.
+// The error is wrapped several layers deep (cleanup.Run -> errors.Join ->
+// NBDProvider.Close -> path flush), all with %w/errors.Join, so errors.Is
+// reaches the underlying syscall.Errno.
+func isExpectedPostCrashFlushError(err error) bool {
+	return errors.Is(err, syscall.EIO) || errors.Is(err, syscall.ENXIO)
+}
+
 // setupSandboxLifecycle starts the goroutine that waits for the sandbox to end
 // and tears it down. releaseTemplate lifts this sandbox's pin on its template;
 // it runs after the sandbox is closed, on every ending (kill, pause, checkpoint
@@ -2465,7 +2477,18 @@ func (s *Server) setupSandboxLifecycle(ctx context.Context, sbx *sandbox.Sandbox
 
 		cleanupErr := sbx.Close(ctx)
 		if cleanupErr != nil {
-			sbxlogger.I(sbx).Error(ctx, "failed to cleanup sandbox, will remove from cache", zap.Error(cleanupErr))
+			if isExpectedPostCrashFlushError(cleanupErr) {
+				// Every ending passes here, including a VM crash / OOM-kill. After
+				// the VM dies the NBD device is in error state, so Close's flush of
+				// /dev/nbdX (fsync + BLKFLSBUF) returns EIO (device error) or ENXIO
+				// (already disconnected). Nothing is lost — the VM already exited —
+				// so this is expected noise, not a cleanup failure; log it at Warn.
+				// Sandbox.Shutdown() (pause/snapshot) flushes through a different
+				// path and still surfaces a genuine EIO as an error.
+				sbxlogger.I(sbx).Warn(ctx, "sandbox device flush failed after VM exit (expected, ignoring)", zap.Error(cleanupErr))
+			} else {
+				sbxlogger.I(sbx).Error(ctx, "failed to cleanup sandbox, will remove from cache", zap.Error(cleanupErr))
+			}
 		}
 
 		closeErr := s.proxy.RemoveFromPool(sbx.LifecycleID)