diff --git a/docs/release-notes/3379.performance.md b/docs/release-notes/3379.performance.md
new file mode 100644
index 0000000000..061c8addcd
--- /dev/null
+++ b/docs/release-notes/3379.performance.md
@@ -0,0 +1 @@
+In `pp.normalize_total`, the median is now computed in-memory when using Dask {smaller}`S Dicks`
diff --git a/src/scanpy/preprocessing/_normalization.py b/src/scanpy/preprocessing/_normalization.py
index e1ee3d4822..fd8fc69ca7 100644
--- a/src/scanpy/preprocessing/_normalization.py
+++ b/src/scanpy/preprocessing/_normalization.py
@@ -32,19 +32,9 @@ def _normalize_data(X, counts, after=None, *, copy: bool = False):
         X = X.astype(np.float32)  # TODO: Check if float64 should be used
     if after is None:
         if isinstance(counts, DaskArray):
-
-            def nonzero_median(x):
-                return np.ma.median(np.ma.masked_array(x, x == 0)).item()
-
-            after = da.from_delayed(
-                dask.delayed(nonzero_median)(counts),
-                shape=(),
-                meta=counts._meta,
-                dtype=counts.dtype,
-            )
-        else:
-            counts_greater_than_zero = counts[counts > 0]
-            after = np.median(counts_greater_than_zero, axis=0)
+            counts = counts.compute()
+        counts_greater_than_zero = counts[counts > 0]
+        after = np.median(counts_greater_than_zero, axis=0)
     counts = counts / after
     return axis_mul_or_truediv(
         X,