diff --git a/packages/db-dtypes/db_dtypes/__init__.py b/packages/db-dtypes/db_dtypes/__init__.py
index bdd8264df5dc..77917fc953b4 100644
--- a/packages/db-dtypes/db_dtypes/__init__.py
+++ b/packages/db-dtypes/db_dtypes/__init__.py
@@ -34,7 +34,7 @@
 date_dtype_name = "dbdate"
 time_dtype_name = "dbtime"
 _EPOCH = datetime.datetime(1970, 1, 1)
-_NPEPOCH = numpy.datetime64(_EPOCH)
+_NPEPOCH = numpy.datetime64(_EPOCH, "ns")
 _NP_DTYPE = "datetime64[ns]"
 
 # Numpy converts datetime64 scalars to datetime.datetime only if microsecond or
@@ -119,7 +119,7 @@ def _datetime(
             )
 
         if pandas.isna(scalar):
-            return numpy.datetime64("NaT")
+            return numpy.datetime64("NaT", "ns")
         if isinstance(scalar, datetime.time):
             return pandas.Timestamp(
                 year=1970,
@@ -172,9 +172,12 @@ def _box_func(self, x):
     __return_deltas = {"timedelta", "timedelta64", "timedelta64[ns]", "<m8", _NP_DTYPE}
 
     def astype(self, dtype, copy=True):
-        deltas = self._ndarray - _NPEPOCH
+        matched_epoch = _NPEPOCH.astype(self._ndarray.dtype)
+        deltas = self._ndarray - matched_epoch
         stype = str(dtype)
         if stype in self.__return_deltas:
+            if "ns" in stype and "ns" not in str(deltas.dtype):
+                return deltas.astype("timedelta64[ns]")
             return deltas
         elif stype.startswith("timedelta64[") or stype.startswith("<m8["):
             return deltas.astype(dtype, copy=False)
@@ -250,7 +253,9 @@ def _datetime(
             scalar = scalar.as_py()
 
         if pandas.isna(scalar):
-            return numpy.datetime64("NaT")
+            # Use day-level "D" resolution for DateArray missing values
+            # to match the underlying storage unit and avoid unit mismatches.
+            return numpy.datetime64("NaT", "D")
         elif isinstance(scalar, numpy.datetime64):
             dateObj = pandas.Timestamp(scalar)
         elif isinstance(scalar, datetime.date):
@@ -321,7 +326,8 @@ def __add__(self, other):
             return self.astype("object") + other
 
         if isinstance(other, TimeArray):
-            return (other._ndarray - _NPEPOCH) + self._ndarray
+            matched_epoch = _NPEPOCH.astype(other._ndarray.dtype)
+            return (other._ndarray - matched_epoch) + self._ndarray
 
         return super().__add__(other)  # type: ignore[misc]
 
diff --git a/packages/db-dtypes/db_dtypes/core.py b/packages/db-dtypes/db_dtypes/core.py
index 6baa46cf2d2d..7f63a474d6e7 100644
--- a/packages/db-dtypes/db_dtypes/core.py
+++ b/packages/db-dtypes/db_dtypes/core.py
@@ -22,6 +22,8 @@
 
 from db_dtypes import pandas_backports
 
+from typing import cast
+
 pandas_release = pandas_backports.pandas_release
 
 
@@ -48,7 +50,16 @@ class BaseDatetimeArray(pandas_backports.OpsMixin, _mixins.NDArrayBackedExtensio
     # Categorical, iNaT for Period. Outside of object dtype, self.isna() should
     # be exactly locations in self._ndarray with _internal_fill_value. See:
     # https://github.com/pandas-dev/pandas/blob/main/pandas/core/arrays/_mixins.py
-    _internal_fill_value = numpy.datetime64("NaT")
+
+    # This is defined dynamically to match the underlying unit of self._ndarray
+    # (e.g., "D" for DateArray, "ns" for TimeArray) to avoid unit mismatches.
+    @property
+    def _internal_fill_value(self):
+        ndarray = getattr(self, "_ndarray", None)
+        if ndarray is not None:
+            unit = cast(Any, numpy.datetime_data(ndarray.dtype)[0])
+            return numpy.datetime64("NaT", unit)
+        return numpy.datetime64("NaT", "ns")
 
     _box_func: Callable[[Any], Any]
     _from_backing_data: Callable[[Any], Any]
diff --git a/packages/db-dtypes/noxfile.py b/packages/db-dtypes/noxfile.py
index b80b1df6832a..2c2c25d03f3e 100644
--- a/packages/db-dtypes/noxfile.py
+++ b/packages/db-dtypes/noxfile.py
@@ -202,7 +202,6 @@ def default(session, tests_path):
         "-W default::FutureWarning",
         f"--junitxml={os.path.split(tests_path)[-1]}_{session.python}_sponge_log.xml",
         "--cov=db_dtypes",
-        "--cov=tests/unit",
         "--cov-append",
         "--cov-config=.coveragerc",
         "--cov-report=",
diff --git a/packages/db-dtypes/tests/unit/test_dtypes.py b/packages/db-dtypes/tests/unit/test_dtypes.py
index 826e0e7517b0..32e246c58fa3 100644
--- a/packages/db-dtypes/tests/unit/test_dtypes.py
+++ b/packages/db-dtypes/tests/unit/test_dtypes.py
@@ -14,6 +14,7 @@
 
 import datetime
 
+import numpy
 import pytest
 
 pd = pytest.importorskip("pandas")
@@ -668,3 +669,21 @@ def test_date_sub():
     do = pd.Series([pd.DateOffset(days=i) for i in range(4)])
     expect = dates.astype("object") - do
     np.testing.assert_array_equal(dates - do, expect)
+
+
+def test_internal_fill_value_without_ndarray():
+    """Ensure fallback return is reached when _ndarray is missing."""
+    from db_dtypes.core import BaseDatetimeArray
+
+    # Create a bare instance using __new__ to bypass __init__ and avoid setting _ndarray
+    instance = BaseDatetimeArray.__new__(BaseDatetimeArray)
+
+    # Extract the value
+    fill_value = instance._internal_fill_value
+
+    # Use numpy.isnat() to check for NaT safely
+    assert numpy.isnat(fill_value)
+
+    # (Optional) Verify the unit metadata is still exactly what you expect
+    unit, _ = numpy.datetime_data(fill_value.dtype)
+    assert unit == "ns"