perf(load): skip Variable.load dispatch for numpy data

doc/whats-new.rst

@@ -42,6 +42,13 @@ Documentation
 Internal Changes
 ~~~~~~~~~~~~~~~~
 
+- Skip the ``to_duck_array`` dispatch in :py:meth:`Variable.load` and
+  :py:meth:`Variable.load_async` when the underlying data is already a
+  ``numpy.ndarray``. The dispatch was a no-op for that case but added
+  noticeable per-variable overhead to :py:meth:`Dataset.load` /
+  :py:meth:`DataArray.load` on datasets with many in-memory variables
+  (:issue:`11352`).
+
 
 .. _whats-new.2026.04.0:
 

xarray/core/variable.py

@@ -1019,6 +1019,11 @@ def load(self, **kwargs) -> Self:
         DataArray.load
         Dataset.load
         """
+        # Fast path: an in-memory numpy ndarray has nothing to load. Subclasses
+        # that advertise a `chunks` attribute (test fakes, third-party chunked
+        # ndarray subclasses) must still go through to_duck_array.
+        if isinstance(self._data, np.ndarray) and not hasattr(self._data, "chunks"):
+            return self
         self._data = to_duck_array(self._data, **kwargs)
         return self
 
@@ -1052,6 +1057,8 @@ async def load_async(self, **kwargs) -> Self:
         DataArray.load_async
         Dataset.load_async
         """
+        if isinstance(self._data, np.ndarray) and not hasattr(self._data, "chunks"):
+            return self
         self._data = await async_to_duck_array(self._data, **kwargs)
         return self