diff --git a/doc/source/whatsnew/v1.4.1.rst b/doc/source/whatsnew/v1.4.1.rst
index b0ad4ce9392f7..3dd0fcdb5b87d 100644
--- a/doc/source/whatsnew/v1.4.1.rst
+++ b/doc/source/whatsnew/v1.4.1.rst
@@ -17,6 +17,7 @@ Fixed regressions
 - Regression in :meth:`Series.mask` with ``inplace=True`` and ``PeriodDtype`` and an incompatible ``other`` coercing to a common dtype instead of raising (:issue:`45546`)
 - Regression in :func:`.assert_frame_equal` not respecting ``check_flags=False`` (:issue:`45554`)
 - Regression in :meth:`Series.fillna` with ``downcast=False`` incorrectly downcasting ``object`` dtype (:issue:`45603`)
+- Regression in :meth:`DataFrame.iat` setting values leading to not propagating correctly in subsequent lookups (:issue:`45684`)
 - Regression when setting values with :meth:`DataFrame.loc` losing :class:`Index` name if :class:`DataFrame` was empty before (:issue:`45621`)
 - Regression in :meth:`~Index.join` with overlapping :class:`IntervalIndex` raising an ``InvalidIndexError`` (:issue:`45661`)
 - Regression in :func:`read_sql` with a DBAPI2 connection that is not an instance of ``sqlite3.Connection`` incorrectly requiring SQLAlchemy be installed (:issue:`45660`)
diff --git a/pandas/conftest.py b/pandas/conftest.py
index 04f460902c11a..3803a10e33fae 100644
--- a/pandas/conftest.py
+++ b/pandas/conftest.py
@@ -1757,6 +1757,14 @@ def indexer_al(request):
     return request.param
 
 
+@pytest.fixture(params=[tm.iat, tm.iloc])
+def indexer_ial(request):
+    """
+    Parametrize over iat.__setitem__, iloc.__setitem__
+    """
+    return request.param
+
+
 @pytest.fixture
 def using_array_manager(request):
     """
diff --git a/pandas/core/frame.py b/pandas/core/frame.py
index 512d85de333ad..a9befb3012343 100644
--- a/pandas/core/frame.py
+++ b/pandas/core/frame.py
@@ -93,7 +93,6 @@
 )
 
 from pandas.core.dtypes.cast import (
-    can_hold_element,
     construct_1d_arraylike_from_scalar,
     construct_2d_arraylike_from_scalar,
     find_common_type,
@@ -3864,23 +3863,16 @@ def _set_value(
         try:
             if takeable:
                 series = self._ixs(col, axis=1)
-                series._set_value(index, value, takeable=True)
-                return
-
-            series = self._get_item_cache(col)
-            loc = self.index.get_loc(index)
-            dtype = series.dtype
-            if isinstance(dtype, np.dtype) and dtype.kind not in ["m", "M"]:
-                # otherwise we have EA values, and this check will be done
-                #  via setitem_inplace
-                if not can_hold_element(series._values, value):
-                    # We'll go through loc and end up casting.
-                    raise TypeError
+                loc = index
+            else:
+                series = self._get_item_cache(col)
+                loc = self.index.get_loc(index)
 
+            # setitem_inplace will do validation that may raise TypeError
+            #  or ValueError
             series._mgr.setitem_inplace(loc, value)
-            # Note: trying to use series._set_value breaks tests in
-            #  tests.frame.indexing.test_indexing and tests.indexing.test_partial
-        except (KeyError, TypeError):
+
+        except (KeyError, TypeError, ValueError):
             # set using a non-recursive method & reset the cache
             if takeable:
                 self.iloc[index, col] = value
diff --git a/pandas/core/internals/base.py b/pandas/core/internals/base.py
index 74d8b20332fff..75ae0f62681aa 100644
--- a/pandas/core/internals/base.py
+++ b/pandas/core/internals/base.py
@@ -9,6 +9,8 @@
     final,
 )
 
+import numpy as np
+
 from pandas._typing import (
     ArrayLike,
     DtypeObj,
@@ -16,7 +18,10 @@
 )
 from pandas.errors import AbstractMethodError
 
-from pandas.core.dtypes.cast import find_common_type
+from pandas.core.dtypes.cast import (
+    find_common_type,
+    np_can_hold_element,
+)
 
 from pandas.core.base import PandasObject
 from pandas.core.indexes.api import (
@@ -174,6 +179,14 @@ def setitem_inplace(self, indexer, value) -> None:
         in place, not returning a new Manager (and Block), and thus never changing
         the dtype.
         """
+        arr = self.array
+
+        # EAs will do this validation in their own __setitem__ methods.
+        if isinstance(arr, np.ndarray):
+            # Note: checking for ndarray instead of np.dtype means we exclude
+            #  dt64/td64, which do their own validation.
+            value = np_can_hold_element(arr.dtype, value)
+
         self.array[indexer] = value
 
     def grouped_reduce(self, func, ignore_failures: bool = False):
diff --git a/pandas/tests/indexing/test_iat.py b/pandas/tests/indexing/test_iat.py
index f1fe464ca0854..44bd51ee1b7d1 100644
--- a/pandas/tests/indexing/test_iat.py
+++ b/pandas/tests/indexing/test_iat.py
@@ -29,3 +29,20 @@ def test_iat_getitem_series_with_period_index():
     expected = ser[index[0]]
     result = ser.iat[0]
     assert expected == result
+
+
+def test_iat_setitem_item_cache_cleared(indexer_ial):
+    # GH#45684
+    data = {"x": np.arange(8, dtype=np.int64), "y": np.int64(0)}
+    df = DataFrame(data).copy()
+    ser = df["y"]
+
+    # previously this iat setting would split the block and fail to clear
+    #  the item_cache.
+    indexer_ial(df)[7, 0] = 9999
+
+    indexer_ial(df)[7, 1] = 1234
+
+    assert df.iat[7, 1] == 1234
+    assert ser.iloc[-1] == 1234
+    assert df.iloc[-1, -1] == 1234