diff --git a/pandas/core/arrays/timedeltas.py b/pandas/core/arrays/timedeltas.py
index 397297c1b88d0..9653121879c0d 100644
--- a/pandas/core/arrays/timedeltas.py
+++ b/pandas/core/arrays/timedeltas.py
@@ -11,7 +11,7 @@
 from pandas import compat
 
 from pandas.core.dtypes.common import (
-    _TD_DTYPE, ensure_int64, is_timedelta64_dtype, is_list_like)
+    _TD_DTYPE, is_list_like)
 from pandas.core.dtypes.generic import ABCSeries
 from pandas.core.dtypes.missing import isna
 
@@ -111,16 +111,16 @@ def dtype(self):
     _attributes = ["freq"]
 
     @classmethod
-    def _simple_new(cls, values, freq=None, **kwargs):
-        values = np.array(values, copy=False)
-        if values.dtype == np.object_:
-            values = array_to_timedelta64(values)
-        if values.dtype != _TD_DTYPE:
-            if is_timedelta64_dtype(values):
-                # non-nano unit
-                values = values.astype(_TD_DTYPE)
-            else:
-                values = ensure_int64(values).view(_TD_DTYPE)
+    def _simple_new(cls, values, freq=None, dtype=_TD_DTYPE):
+        # `dtype` is passed by _shallow_copy in corner cases, should always
+        #  be timedelta64[ns] if present
+        assert dtype == _TD_DTYPE
+        assert isinstance(values, np.ndarray), type(values)
+
+        if values.dtype == 'i8':
+            values = values.view('m8[ns]')
+
+        assert values.dtype == 'm8[ns]'
 
         result = object.__new__(cls)
         result._data = values
@@ -131,6 +131,10 @@ def __new__(cls, values, freq=None):
 
         freq, freq_infer = dtl.maybe_infer_freq(freq)
 
+        values = np.array(values, copy=False)
+        if values.dtype == np.object_:
+            values = array_to_timedelta64(values)
+
         result = cls._simple_new(values, freq=freq)
         if freq_infer:
             inferred = result.inferred_freq
@@ -166,17 +170,15 @@ def _generate_range(cls, start, end, periods, freq, closed=None):
 
         if freq is not None:
             index = _generate_regular_range(start, end, periods, freq)
-            index = cls._simple_new(index, freq=freq)
         else:
             index = np.linspace(start.value, end.value, periods).astype('i8')
-            index = cls._simple_new(index, freq=freq)
 
         if not left_closed:
             index = index[1:]
         if not right_closed:
             index = index[:-1]
 
-        return index
+        return cls._simple_new(index, freq=freq)
 
     # ----------------------------------------------------------------
     # Arithmetic Methods
diff --git a/pandas/core/indexes/timedeltas.py b/pandas/core/indexes/timedeltas.py
index e5da21478d0a4..22ecefae8cbe2 100644
--- a/pandas/core/indexes/timedeltas.py
+++ b/pandas/core/indexes/timedeltas.py
@@ -35,6 +35,7 @@
     to_timedelta, _coerce_scalar_to_timedelta_type)
 from pandas._libs import (lib, index as libindex,
                           join as libjoin, Timedelta, NaT)
+from pandas._libs.tslibs.timedeltas import array_to_timedelta64
 
 
 class TimedeltaIndex(TimedeltaArrayMixin, DatetimeIndexOpsMixin,
@@ -166,6 +167,19 @@ def __new__(cls, data=None, unit=None, freq=None, start=None, end=None,
         elif copy:
             data = np.array(data, copy=True)
 
+        data = np.array(data, copy=False)
+        if data.dtype == np.object_:
+            data = array_to_timedelta64(data)
+        if data.dtype != _TD_DTYPE:
+            if is_timedelta64_dtype(data):
+                # non-nano unit
+                # TODO: watch out for overflows
+                data = data.astype(_TD_DTYPE)
+            else:
+                data = ensure_int64(data).view(_TD_DTYPE)
+
+        assert data.dtype == 'm8[ns]', data.dtype
+
         subarr = cls._simple_new(data, name=name, freq=freq)
         # check that we are matching freqs
         if verify_integrity and len(subarr) > 0:
@@ -180,12 +194,23 @@ def __new__(cls, data=None, unit=None, freq=None, start=None, end=None,
         return subarr
 
     @classmethod
-    def _simple_new(cls, values, name=None, freq=None, **kwargs):
-        result = super(TimedeltaIndex, cls)._simple_new(values, freq, **kwargs)
+    def _simple_new(cls, values, name=None, freq=None, dtype=_TD_DTYPE):
+        # `dtype` is passed by _shallow_copy in corner cases, should always
+        #  be timedelta64[ns] if present
+        assert dtype == _TD_DTYPE
+
+        assert isinstance(values, np.ndarray), type(values)
+        if values.dtype == 'i8':
+            values = values.view('m8[ns]')
+        assert values.dtype == 'm8[ns]', values.dtype
+
+        result = super(TimedeltaIndex, cls)._simple_new(values, freq)
         result.name = name
         result._reset_identity()
         return result
 
+    _shallow_copy = Index._shallow_copy
+
     @property
     def _formatter_func(self):
         from pandas.io.formats.format import _get_format_timedelta64