diff --git a/lib/iris/experimental/xarray_dataset_wrapper.py b/lib/iris/experimental/xarray_dataset_wrapper.py
index 455057afad..c453adfcff 100644
--- a/lib/iris/experimental/xarray_dataset_wrapper.py
+++ b/lib/iris/experimental/xarray_dataset_wrapper.py
@@ -14,9 +14,10 @@
 However, this is a convenient place to test, for now.
 
 """
-from collections import OrderedDict
+from collections import Iterable, OrderedDict
 from typing import Optional
 
+import dask.array as da
 import netCDF4 as nc
 import numpy as np
 import xarray
@@ -177,8 +178,74 @@ def size(self):
     #
     # writing
     #
+
+    # This identifies a VariableMimic as an array-like object to which LAZY
+    # data can be written (a dask array).
+    _ACCEPTS_LAZY_DATA_WRITES = True
+    # Iris netcdf saving can recognise this, and will "save" the lazy content
+    # into a VariableMimic, instead of streaming real data into it as it would
+    # to a normal file variable.
+    # Xarray itself understands dask content, and will stream it into a file.
+
     def __setitem__(self, keys, data):
-        self._xr[keys] = data
+        # Note: assigned 'data' may be real or lazy.
+        # Lazy input can *stay* lazy, in at least some cases, since xarray can
+        # handle lazy content.
+        if self._indexes_address_full_shape(keys):
+            # This assignment replaces the entire array content.  So, we can
+            # just write it into the xarray variable.data : this will preserve
+            # laziness, since xarray supports lazy data content.
+            self._xr.data = data
+        else:
+            if hasattr(self._xr.data, "compute"):
+                # Attempted *partial* overwrite of lazy content.
+                # Can't be done without losing the laziness (or at least, is
+                # not implemented at present).
+                # Fetch the real array content, so we can overwrite part of it.
+                self._xr.data = self._xr.data.compute()
+            if hasattr(data, "compute"):
+                # Assigned value is lazy : fetch real data.
+                data = data.compute()
+            # rewrite the selected portion of the array content.
+            self._xr[keys] = data
+
+    def _indexes_address_full_shape(self, keys):
+        """Check that a set of indexing keys addresses an entire array."""
+        shape = self.shape
+        if not isinstance(keys, Iterable):
+            keys = (keys,)
+
+        # First check there aren't more keys than dims.
+        # N.B. this means we don't support np.newaxis / None
+        # N.B. *fewer* keys is ok, as remainder is then equivalent to ", ..."
+        ok = len(keys) <= len(shape)
+
+        # Remove final Ellipses (which have no effect) and error non-final ones
+        if ok:
+            while keys[-1] is Ellipsis:
+                keys = keys[:-1]
+
+        # What remains can *only* be slice(a,b) with suitable a+b,
+        # *or* 0 for a length-1 dim (nothing else being "full")
+        if ok:
+            for key, dimlen in zip(keys, shape):
+                if key == 0:
+                    ok = dimlen == 1
+                elif isinstance(key, slice):
+                    start, stop, step = key.start, key.stop, key.step
+                    if step not in (1, None):
+                        ok = False
+                    elif start not in (0, None):
+                        ok = False
+                    elif stop is not None and stop < dimlen:
+                        ok = False
+                else:
+                    # We don't support *any* other types,
+                    # e.g. None / dropaxis / newaxis / non-final Ellipsis
+                    ok = False
+                if not ok:
+                    break
+        return ok
 
 
 class DatasetMimic(_Nc4AttrsMimic):
@@ -364,7 +431,10 @@ def createVariable(
             attrs = {}
             dt_code = f"{datatype.kind}{datatype.itemsize}"
             use_fill = nc.default_fillvals[dt_code]
-        data = np.full(shape, fill_value=use_fill, dtype=datatype)
+        # Use a DASK array : this is perfectly reasonable and won't chew up a
+        # lot of space.  This content is usually discarded it when assigning
+        # variable data afterwards, anywway.
+        data = da.full(shape, fill_value=use_fill, dtype=datatype)
 
         xr_var = xr.Variable(dims=dimensions, data=data, attrs=attrs)
         original_varname = varname
diff --git a/lib/iris/fileformats/netcdf.py b/lib/iris/fileformats/netcdf.py
index 4efed43db9..73dd857ded 100644
--- a/lib/iris/fileformats/netcdf.py
+++ b/lib/iris/fileformats/netcdf.py
@@ -3017,12 +3017,27 @@ def _lazy_stream_data(data, fill_value, fill_warn, cf_var):
 
         if is_lazy_data(data):
 
-            def store(data, cf_var, fill_value):
-                # Store lazy data and check whether it is masked and contains
-                # the fill value
-                target = _FillValueMaskCheckAndStoreTarget(cf_var, fill_value)
-                da.store([data], [target])
-                return target.is_masked, target.contains_value
+            if hasattr(cf_var, "_ACCEPTS_LAZY_DATA_WRITES"):
+                # A special case : we can write a lazy array directly into the
+                # the output target.
+                # In this case, instead of performing an actual streaming
+                # write, we simply assign a lazy array into the variable.
+                # CAVEAT: this does not actually write the file, nor perform
+                # the expected fill-value check ...
+                def store(data, cf_var, fill_value):
+                    cf_var[:] = data
+                    return False, False
+
+            else:
+
+                def store(data, cf_var, fill_value):
+                    # Store lazy data and check whether it is masked and contains
+                    # the fill value
+                    target = _FillValueMaskCheckAndStoreTarget(
+                        cf_var, fill_value
+                    )
+                    da.store([data], [target])
+                    return target.is_masked, target.contains_value
 
         else:
 
diff --git a/lib/iris/tests/integration/experimental/test_xarray_dataset_wrapper.py b/lib/iris/tests/integration/experimental/test_xarray_dataset_wrapper.py
index 48d9878007..fd41c79c69 100644
--- a/lib/iris/tests/integration/experimental/test_xarray_dataset_wrapper.py
+++ b/lib/iris/tests/integration/experimental/test_xarray_dataset_wrapper.py
@@ -116,6 +116,16 @@ def test_1(self):
         iris.save(cubes, nc_faked_xr, saver="nc")
         ds = nc_faked_xr.to_xarray_dataset()
 
+        # Special "lazy streaming" ensures that the main cube arrays are LAZY.
+        for cube in cubes:
+            assert hasattr(ds.variables[cube.var_name].data, "compute")
+
+        # print('')
+        # for varname, var in ds.variables.items():
+        #     print(f'Variable {varname!s} : shape {var.shape}, type {type(var.data)}')
+        # print('')
+
+        # Save the netcdf version.
         xr_outpath = str(Path("tmp_xr.nc").absolute())
         ds.to_netcdf(xr_outpath)
 
@@ -143,3 +153,10 @@ def capture_dump_lines(filepath_str):
         # save, or via iris.save --> xarray.Dataset --> xarray.Data.to_netcdf()
         # Compare, omitting the first line with the filename
         assert lines_xr_save[1:] == lines_iris_save[1:]
+
+        # Also check that the *content* is identical.
+        iris_saved_reload = iris.load("tmp_iris.nc")
+        xr_saved_reload = iris.load("tmp_xr.nc")
+        iris_saved_reload = sorted(iris_saved_reload, key=lambda c: c.name())
+        xr_saved_reload = sorted(xr_saved_reload, key=lambda c: c.name())
+        assert xr_saved_reload == iris_saved_reload