diff --git a/packages/orchestrator/pkg/sandbox/nbd/pool.go b/packages/orchestrator/pkg/sandbox/nbd/pool.go
index ba6e3e1198..2d11ff561c 100644
--- a/packages/orchestrator/pkg/sandbox/nbd/pool.go
+++ b/packages/orchestrator/pkg/sandbox/nbd/pool.go
@@ -26,6 +26,9 @@ const (
 	waitOnNBDError                = 50 * time.Millisecond
 	devicePoolCloseReleaseTimeout = 10 * time.Minute
 	sysBlockDir                   = "/sys/block"
+	// releaseRetryDelay is how long ReleaseDevice waits before retrying a device
+	// that is still in use.
+	releaseRetryDelay = 500 * time.Millisecond
 )
 
 var (
@@ -366,7 +369,14 @@ func (d *DevicePool) ReleaseDevice(ctx context.Context, idx DeviceSlot, opts ...
 			logger.L().Error(ctx, "error releasing device", zap.Int("attempt", attempt), zap.Error(err))
 		}
 
-		time.Sleep(500 * time.Millisecond)
+		// Wait on the context too: Close bounds a stuck device with WithTimeout,
+		// and shutdown cancels the context outright. An uninterruptible sleep
+		// here would make every remaining attempt outlive both.
+		select {
+		case <-ctx.Done():
+			return ctx.Err()
+		case <-time.After(releaseRetryDelay):
+		}
 	}
 }
 
diff --git a/packages/orchestrator/pkg/sandbox/nbd/pool_release_test.go b/packages/orchestrator/pkg/sandbox/nbd/pool_release_test.go
new file mode 100644
index 0000000000..114de95680
--- /dev/null
+++ b/packages/orchestrator/pkg/sandbox/nbd/pool_release_test.go
@@ -0,0 +1,82 @@
+//go:build linux
+
+package nbd
+
+import (
+	"context"
+	"testing"
+	"time"
+
+	"github.com/bits-and-blooms/bitset"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+// unreachableSlot is an index no NBD device can have, so isDeviceFree always
+// fails to read /sys/block/nbd<idx>/size and release() keeps returning an
+// error. That drives the retry loop deterministically, whether or not the nbd
+// module is loaded on the machine running the test.
+const unreachableSlot = DeviceSlot(1 << 20)
+
+func retryingPool() *DevicePool {
+	return &DevicePool{
+		done:      make(chan struct{}),
+		usedSlots: bitset.New(16),
+		slots:     make(chan DeviceSlot, 1),
+	}
+}
+
+// A cancelled context must abort the release retry loop while it is backing
+// off, not only when it happens to be at the top of the loop. Otherwise every
+// stuck device adds a full backoff interval to orchestrator shutdown.
+func TestReleaseDeviceAbortsBackoffOnContextCancel(t *testing.T) {
+	t.Parallel()
+
+	pool := retryingPool()
+
+	ctx, cancel := context.WithCancel(t.Context())
+	defer cancel()
+
+	// Cancel once the first failed attempt has entered the backoff.
+	time.AfterFunc(releaseRetryDelay/10, cancel)
+
+	start := time.Now()
+	err := pool.ReleaseDevice(ctx, unreachableSlot, WithInfiniteRetry())
+	elapsed := time.Since(start)
+
+	require.ErrorIs(t, err, context.Canceled)
+	assert.Less(t, elapsed, releaseRetryDelay,
+		"ReleaseDevice slept through the cancellation instead of aborting the backoff")
+}
+
+// The same applies to the deadline installed by WithTimeout, which Close relies
+// on to bound how long a single stuck device can hold up the pool.
+func TestReleaseDeviceAbortsBackoffOnTimeout(t *testing.T) {
+	t.Parallel()
+
+	pool := retryingPool()
+
+	start := time.Now()
+	err := pool.ReleaseDevice(t.Context(), unreachableSlot,
+		WithInfiniteRetry(),
+		WithTimeout(releaseRetryDelay/10),
+	)
+	elapsed := time.Since(start)
+
+	require.ErrorIs(t, err, context.DeadlineExceeded)
+	assert.Less(t, elapsed, releaseRetryDelay,
+		"ReleaseDevice slept past its own deadline")
+}
+
+// Without infinite retry the first failure is returned immediately, so the
+// backoff change must not alter that path.
+func TestReleaseDeviceReturnsFirstErrorWithoutInfiniteRetry(t *testing.T) {
+	t.Parallel()
+
+	pool := retryingPool()
+
+	err := pool.ReleaseDevice(t.Context(), unreachableSlot)
+
+	require.Error(t, err)
+	assert.NotErrorIs(t, err, context.Canceled)
+}