diff --git a/pyproject.toml b/pyproject.toml
index ce52ddf34a..d7d01375ef 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -182,6 +182,7 @@ disable = [
     "import-outside-toplevel",
     "inconsistent-return-statements",
     "invalid-class-object",
+    "invalid-metaclass",
     "invalid-name",
     "invalid-unary-operand-type",
     "keyword-arg-before-vararg",
diff --git a/src/awkward/_backends/backend.py b/src/awkward/_backends/backend.py
index 7763953ed1..a7da8102bb 100644
--- a/src/awkward/_backends/backend.py
+++ b/src/awkward/_backends/backend.py
@@ -10,7 +10,7 @@
 from awkward._nplikes.numpy import Numpy
 from awkward._nplikes.numpy_like import NumpyLike, NumpyMetadata
 from awkward._singleton import PublicSingleton
-from awkward._typing import Callable, TypeAlias, TypeVar
+from awkward._typing import Callable, NominalMeta, TypeAlias, TypeVar
 
 np = NumpyMetadata.instance()
 numpy = Numpy.instance()
@@ -21,7 +21,7 @@
 KernelType: TypeAlias = "Callable[..., KernelError | None]"
 
 
-class Backend(PublicSingleton, ABC):
+class Backend(PublicSingleton, ABC, metaclass=NominalMeta):
     name: str
 
     @property
diff --git a/src/awkward/_nplikes/array_like.py b/src/awkward/_nplikes/array_like.py
index 609ae84aeb..857350faee 100644
--- a/src/awkward/_nplikes/array_like.py
+++ b/src/awkward/_nplikes/array_like.py
@@ -10,6 +10,7 @@
     Any,
     DType,
     EllipsisType,
+    NominalMeta,
     Protocol,
     Self,
     SupportsIndex,
@@ -140,7 +141,7 @@ def __or__(self, other: int | bool | Self) -> Self: ...
     def __invert__(self) -> Self: ...
 
 
-class MaterializableArray(ArrayLike):
+class MaterializableArray(ArrayLike, metaclass=NominalMeta):
     @abstractmethod
     def materialize(self) -> ArrayLike: ...
 
diff --git a/src/awkward/_nplikes/array_module.py b/src/awkward/_nplikes/array_module.py
index a6eae53195..ba19813dda 100644
--- a/src/awkward/_nplikes/array_module.py
+++ b/src/awkward/_nplikes/array_module.py
@@ -18,7 +18,16 @@
 from awkward._nplikes.placeholder import PlaceholderArray
 from awkward._nplikes.shape import ShapeItem, unknown_length
 from awkward._nplikes.virtual import VirtualNDArray
-from awkward._typing import TYPE_CHECKING, Any, DType, Final, Literal, TypeVar, cast
+from awkward._typing import (
+    TYPE_CHECKING,
+    Any,
+    DType,
+    Final,
+    Literal,
+    NominalMeta,
+    TypeVar,
+    cast,
+)
 
 if TYPE_CHECKING:
     from numpy.typing import DTypeLike
@@ -56,7 +65,7 @@ def _nplike_unique_has_equal_nan(module: Any) -> bool:
 ArrayLikeT = TypeVar("ArrayLikeT", bound=ArrayLike)
 
 
-class ArrayModuleNumpyLike(NumpyLike[ArrayLikeT]):
+class ArrayModuleNumpyLike(NumpyLike[ArrayLikeT], metaclass=NominalMeta):
     """
     Some methods maintain virtualness while others are required to materialize the array.
     The `maybe_materialize` function is used for that purpose.
diff --git a/src/awkward/_nplikes/typetracer.py b/src/awkward/_nplikes/typetracer.py
index 792324872e..cfb0855887 100644
--- a/src/awkward/_nplikes/typetracer.py
+++ b/src/awkward/_nplikes/typetracer.py
@@ -29,6 +29,7 @@
     EllipsisType,
     Final,
     Literal,
+    NominalMeta,
     Self,
     SupportsIndex,
     TypeGuard,
@@ -236,7 +237,7 @@ def data_touched_in(self, node_ids: Collection[str]) -> list[str]:
         return list(out)
 
 
-class TypeTracerArray(NDArrayOperatorsMixin, ArrayLike):
+class TypeTracerArray(NDArrayOperatorsMixin, ArrayLike, metaclass=NominalMeta):
     _dtype: numpy.dtype
     _shape: tuple[ShapeItem, ...]
 
@@ -602,7 +603,7 @@ def try_touch_shape(array: Any):
 
 
 @register_nplike
-class TypeTracer(NumpyLike[TypeTracerArray]):
+class TypeTracer(NumpyLike[TypeTracerArray], metaclass=NominalMeta):
     known_data: Final = False
     is_eager: Final = True
     supports_structured_dtypes: Final = True
diff --git a/src/awkward/_typing.py b/src/awkward/_typing.py
index 230a83e2cb..b5336baa53 100644
--- a/src/awkward/_typing.py
+++ b/src/awkward/_typing.py
@@ -79,3 +79,18 @@
 JSONMapping: TypeAlias = "dict[str, JSONSerializable]"
 
 DType: TypeAlias = numpy.dtype
+
+
+class NominalMeta(type(Protocol)):
+    """Metaclass restoring the interpreter's ``isinstance`` fast path.
+
+    Concrete classes that descend from a ``Protocol`` inherit
+    ``typing._ProtocolMeta``, whose Python-level ``__instancecheck__`` costs
+    roughly four times a plain class check even though, for a non-protocol
+    subclass, it only performs the ordinary nominal test. Classes using this
+    metaclass are checked nominally and cannot be given virtual subclasses
+    through ``abc``'s ``register()``.
+    """
+
+    __instancecheck__ = type.__instancecheck__
+    __subclasscheck__ = type.__subclasscheck__
diff --git a/tests/test_4339_nominal_isinstance.py b/tests/test_4339_nominal_isinstance.py
new file mode 100644
index 0000000000..2d93a7932a
--- /dev/null
+++ b/tests/test_4339_nominal_isinstance.py
@@ -0,0 +1,121 @@
+# BSD 3-Clause License; see https://github.com/scikit-hep/awkward/blob/main/LICENSE
+
+from __future__ import annotations
+
+import numpy as np
+import pytest
+
+import awkward as ak
+from awkward._backends.backend import Backend
+from awkward._backends.numpy import NumpyBackend
+from awkward._backends.typetracer import TypeTracerBackend
+from awkward._nplikes.array_like import ArrayLike, MaterializableArray
+from awkward._nplikes.array_module import ArrayModuleNumpyLike
+from awkward._nplikes.numpy import Numpy
+from awkward._nplikes.placeholder import PlaceholderArray
+from awkward._nplikes.shape import unknown_length
+from awkward._nplikes.typetracer import TypeTracer, TypeTracerArray
+from awkward._nplikes.virtual import VirtualNDArray
+from awkward._typing import NominalMeta
+
+# Every class that opts into the nominal metaclass.
+NOMINAL_CLASSES = [
+    Backend,
+    MaterializableArray,
+    ArrayModuleNumpyLike,
+    TypeTracer,
+    TypeTracerArray,
+]
+
+
+@pytest.mark.parametrize("cls", NOMINAL_CLASSES, ids=lambda cls: cls.__name__)
+def test_uses_nominal_metaclass(cls):
+    assert type(cls) is NominalMeta
+
+
+def test_real_implementations_are_still_instances():
+    # nplikes
+    assert isinstance(Numpy.instance(), ArrayModuleNumpyLike)
+    assert issubclass(Numpy, ArrayModuleNumpyLike)
+    assert isinstance(TypeTracer.instance(), TypeTracer)
+
+    # backends
+    assert isinstance(NumpyBackend.instance(), Backend)
+    assert isinstance(TypeTracerBackend.instance(), Backend)
+
+    # arrays
+    placeholder = PlaceholderArray(Numpy.instance(), (3,), np.dtype(np.int64))
+    assert isinstance(placeholder, MaterializableArray)
+    virtual = VirtualNDArray(
+        Numpy.instance(), (3,), np.dtype(np.int64), lambda: np.arange(3)
+    )
+    assert isinstance(virtual, MaterializableArray)
+
+    layout = ak.Array([[1, 2, 3], [], [4, 5]]).layout.to_typetracer(forget_length=True)
+    assert isinstance(layout.content.data, TypeTracerArray)
+    assert isinstance(layout.backend, Backend)
+    assert isinstance(layout.backend.nplike, TypeTracer)
+
+    # the ArrayLike protocol is still part of the concrete classes' interface
+    assert ArrayLike in MaterializableArray.__mro__
+    assert ArrayLike in TypeTracerArray.__mro__
+
+
+@pytest.mark.parametrize("cls", NOMINAL_CLASSES, ids=lambda cls: cls.__name__)
+def test_duck_types_are_not_instances(cls):
+    """A structurally compatible duck type is *not* an instance.
+
+    This is the behaviour change: these classes are checked nominally, so
+    neither structural compatibility nor an explicit ``abc`` ``register()``
+    makes an unrelated class an instance. Only real subclasses count.
+    """
+
+    class Duck:
+        known_data = True
+        dtype = np.dtype(np.int64)
+        ndim = 1
+        shape = (3,)
+        strides = (8,)
+        name = "duck"
+
+        def materialize(self):
+            return np.arange(3)
+
+    assert not isinstance(Duck(), cls)
+    cls.register(Duck)
+    assert not isinstance(Duck(), cls)
+    assert not issubclass(Duck, cls)
+
+
+def test_explicit_subclasses_are_still_instances():
+    class MyPlaceholder(PlaceholderArray):
+        pass
+
+    assert isinstance(
+        MyPlaceholder(Numpy.instance(), (unknown_length,), np.dtype(np.int64)),
+        MaterializableArray,
+    )
+
+    class MyBackend(NumpyBackend):
+        pass
+
+    assert issubclass(MyBackend, Backend)
+    assert isinstance(MyBackend.instance(), Backend)
+
+    class MyNumpy(Numpy):
+        pass
+
+    assert issubclass(MyNumpy, ArrayModuleNumpyLike)
+
+
+def test_operations_are_unaffected():
+    array = ak.Array([[1, 2, 3], [], [4, 5]])
+    assert ak.num(array).to_list() == [3, 0, 2]
+    assert (array * 2).to_list() == [[2, 4, 6], [], [8, 10]]
+    assert array[1:].to_list() == [[], [4, 5]]
+    assert ak.sum(array, axis=-1).to_list() == [6, 0, 9]
+
+    tracer = ak.Array(array.layout.to_typetracer(forget_length=True))
+    assert ak.num(tracer).layout.dtype == np.dtype(np.int64)
+    assert (tracer * 2).layout.form == (array * 2).layout.form
+    assert ak.sum(tracer, axis=-1).layout.form == ak.sum(array, axis=-1).layout.form