diff --git a/docs/src/whatsnew/latest.rst b/docs/src/whatsnew/latest.rst
index 68737d86cc..0c9c58a203 100644
--- a/docs/src/whatsnew/latest.rst
+++ b/docs/src/whatsnew/latest.rst
@@ -61,7 +61,7 @@ This document explains the changes made to Iris for this release
 ===========================
 
 #. `@bouweandela`_ made loading :class:`~iris.cube.Cube`s from NetCDF files
-   faster. (:pull:`6229` and :pull:`6252`)
+   faster. (:pull:`6229`, :pull:`6252`, :pull:`6231`)
 
 #. `@fnattino`_ enabled lazy cube interpolation using the linear and
    nearest-neighbour interpolators (:class:`iris.analysis.Linear` and
diff --git a/lib/iris/fileformats/netcdf/_thread_safe_nc.py b/lib/iris/fileformats/netcdf/_thread_safe_nc.py
index 3a556f5447..9956f1480a 100644
--- a/lib/iris/fileformats/netcdf/_thread_safe_nc.py
+++ b/lib/iris/fileformats/netcdf/_thread_safe_nc.py
@@ -327,6 +327,12 @@ def ndim(self):
     def dask_meta(self):
         return np.ma.array(np.empty((0,) * self.ndim, dtype=self.dtype), mask=True)
 
+    def __dask_tokenize__(self):
+        # Dask uses this function to uniquely identify the "array".
+        # A custom function is slightly faster than general object tokenization,
+        # which improves the speed of loading small NetCDF files.
+        return f"<{self.__class__.__name__} path={self.path!r} variable_name={self.variable_name!r}>"
+
     def __getitem__(self, keys):
         # Using a DatasetWrapper causes problems with invalid ID's and the
         # netCDF4 library, presumably because __getitem__ gets called so many