diff --git a/dask_expr/_collection.py b/dask_expr/_collection.py
index b317efb9a..002aff4da 100644
--- a/dask_expr/_collection.py
+++ b/dask_expr/_collection.py
@@ -1571,6 +1571,8 @@ def std(
                     )
                 else:
                     needs_time_conversion = True
+                    if axis == 1:
+                        numeric_dd = numeric_dd.astype(f"datetime64[{meta.array.unit}]")
                     for col in time_cols:
                         numeric_dd[col] = _convert_to_numeric(numeric_dd[col], skipna)
         else:
@@ -1583,6 +1585,11 @@ def std(
             units = [getattr(self._meta[c].array, "unit", None) for c in time_cols]
 
         if axis == 1:
+            _kwargs = (
+                {}
+                if not needs_time_conversion
+                else {"unit": meta.array.unit, "dtype": meta.dtype}
+            )
             return numeric_dd.map_partitions(
                 M.std if not needs_time_conversion else _sqrt_and_convert_to_timedelta,
                 meta=meta,
@@ -1591,6 +1598,7 @@ def std(
                 ddof=ddof,
                 enforce_metadata=False,
                 numeric_only=numeric_only,
+                **_kwargs,
             )
 
         result = numeric_dd.var(
diff --git a/dask_expr/_describe.py b/dask_expr/_describe.py
index 5df15a3a7..a1878b4aa 100644
--- a/dask_expr/_describe.py
+++ b/dask_expr/_describe.py
@@ -52,21 +52,25 @@ def _lower(self):
             frame.max(split_every=self.split_every),
         ]
         return DescribeNumericAggregate(
-            self.frame._meta.name, is_td_col, is_dt_col, *stats
+            self.frame._meta.name,
+            is_td_col,
+            is_dt_col,
+            getattr(self.frame._meta.array, "unit", None),
+            *stats,
         )
 
 
 class DescribeNumericAggregate(Blockwise):
-    _parameters = ["name", "is_timedelta_col", "is_datetime_col"]
+    _parameters = ["name", "is_timedelta_col", "is_datetime_col", "unit"]
     _defaults = {"is_timedelta_col": False, "is_datetime_col": False}
 
     def _broadcast_dep(self, dep):
         return dep.npartitions == 1
 
     @staticmethod
-    def operation(name, is_timedelta_col, is_datetime_col, *stats):
+    def operation(name, is_timedelta_col, is_datetime_col, unit, *stats):
         return describe_numeric_aggregate(
-            stats, name, is_timedelta_col, is_datetime_col
+            stats, name, is_timedelta_col, is_datetime_col, unit
         )