BUG: timedelta64(NaT) incorrectly treated as datetime in some dataframe ops

pandas/core/ops/__init__.py

@@ -499,8 +499,19 @@ def column_op(a, b):
         # in which case we specifically want to operate row-by-row
         assert right.index.equals(left.columns)
 
-        def column_op(a, b):
-            return {i: func(a.iloc[:, i], b.iloc[i]) for i in range(len(a.columns))}
+        if right.dtype == "timedelta64[ns]":
+            # ensure we treat NaT values as the correct dtype
+            # Note: we do not do this unconditionally as it may be lossy or
+            #  expensive for EA dtypes.
+            right = np.asarray(right)
+
+            def column_op(a, b):
+                return {i: func(a.iloc[:, i], b[i]) for i in range(len(a.columns))}
+
+        else:
+
+            def column_op(a, b):
+                return {i: func(a.iloc[:, i], b.iloc[i]) for i in range(len(a.columns))}
 
     elif isinstance(right, ABCSeries):
         assert right.index.equals(left.index)  # Handle other cases later

pandas/tests/frame/test_arithmetic.py

@@ -457,6 +457,16 @@ def test_arith_flex_zero_len_raises(self):
 
 
 class TestFrameArithmetic:
+    def test_td64_op_nat_casting(self):
+        # Make sure we don't accidentally treat timedelta64(NaT) as datetime64
+        #  when calling dispatch_to_series in DataFrame arithmetic
+        ser = pd.Series(["NaT", "NaT"], dtype="timedelta64[ns]")
+        df = pd.DataFrame([[1, 2], [3, 4]])
+
+        result = df * ser
+        expected = pd.DataFrame({0: ser, 1: ser})
+        tm.assert_frame_equal(result, expected)
+
     def test_df_add_2d_array_rowlike_broadcasts(self):
         # GH#23000
         arr = np.arange(6).reshape(3, 2)