diff --git a/xrspatial/geotiff/_backends/gpu.py b/xrspatial/geotiff/_backends/gpu.py
index c84ea30b..616a27f7 100644
--- a/xrspatial/geotiff/_backends/gpu.py
+++ b/xrspatial/geotiff/_backends/gpu.py
@@ -210,10 +210,13 @@ def read_geotiff_gpu(source: str, *,
     stable_only : bool, default False
         [experimental] Read-side opt-in for stable-tier sources only.
         The GPU read path does not consume VRT sources directly (VRT
-        routing happens in ``open_geotiff``), so this kwarg is accepted
-        for cross-backend signature symmetry and is a no-op on the GPU
-        eager / chunked paths. See ``open_geotiff`` for the full
-        description.
+        routing happens in ``open_geotiff``), but it does route HTTP /
+        fsspec sources through the CPU fallback, and those advanced-tier
+        readers must be gated. With ``stable_only=True`` a remote source
+        raises ``RemoteStableSourcesOnlyError`` before any cupy import or
+        decode, matching ``open_geotiff`` and ``read_geotiff_dask``. Pass
+        ``allow_experimental_codecs=True`` to unlock the advanced tier.
+        See ``open_geotiff`` for the full description.
     allow_experimental_codecs : bool, default False
         [experimental] Read-side opt-in for Tier 3 experimental codecs
         (``lerc``, ``jpeg2000`` / ``j2k``, ``lz4``). The GPU read path
@@ -275,6 +278,21 @@ def read_geotiff_gpu(source: str, *,
         max_cloud_bytes=max_cloud_bytes,
     )
 
+    # ``open_geotiff`` and ``read_geotiff_dask`` gate ``stable_only=True``
+    # for advanced-tier HTTP / fsspec sources before dispatching. This GPU
+    # entry point is also a direct public reader and it routes remote
+    # sources through the CPU fallback below, so a direct caller could read
+    # an advanced-tier remote source under ``stable_only=True`` unless the
+    # same gate fires here. Run it before the cupy import, the CUDA
+    # preflight, and the remote fallback so the rejection is independent of
+    # GPU availability, matching the other two readers.
+    from .._validation import _validate_stable_only_remote
+    _validate_stable_only_remote(
+        source,
+        stable_only=stable_only,
+        allow_experimental_codecs=allow_experimental_codecs,
+    )
+
     new_passed = on_gpu_failure is not _ON_GPU_FAILURE_SENTINEL
     old_passed = gpu is not _GPU_DEPRECATED_SENTINEL
     if new_passed and old_passed:
diff --git a/xrspatial/geotiff/tests/test_stable_only_remote_2821.py b/xrspatial/geotiff/tests/test_stable_only_remote_2821.py
index 32ba6ef2..09eb645f 100644
--- a/xrspatial/geotiff/tests/test_stable_only_remote_2821.py
+++ b/xrspatial/geotiff/tests/test_stable_only_remote_2821.py
@@ -22,12 +22,24 @@
 import pytest
 
 from xrspatial.geotiff import (GeoTIFFAmbiguousMetadataError, RemoteStableSourcesOnlyError,
-                               open_geotiff, read_geotiff_dask)
+                               open_geotiff, read_geotiff_dask, read_geotiff_gpu)
 from xrspatial.geotiff.tests._helpers.tiff_builders import make_minimal_tiff
 
 fsspec = pytest.importorskip("fsspec")
 
 
+def _gpu_available() -> bool:
+    try:
+        import cupy
+        return bool(cupy.cuda.is_available())
+    except Exception:
+        return False
+
+
+_HAS_GPU = _gpu_available()
+_gpu_only = pytest.mark.skipif(not _HAS_GPU, reason="cupy + CUDA required")
+
+
 _MEMORY_URL = "memory:///stable_only_2821/sample.tif"
 
 
@@ -134,3 +146,38 @@ def test_local_dask_source_succeeds_under_stable_only(local_tiff_path):
     """A plain local-file dask read still works under ``stable_only=True``."""
     result = open_geotiff(local_tiff_path, stable_only=True, chunks=2)
     assert result.shape == (4, 4)
+
+
+# ---------------------------------------------------------------------------
+# Direct GPU entry point (issue #2867)
+#
+# ``read_geotiff_gpu`` is a public direct reader that routes HTTP / fsspec
+# sources through the CPU fallback, so it must apply the same remote gate as
+# ``open_geotiff`` and ``read_geotiff_dask``. The gate runs before the cupy
+# import and the CUDA preflight, so the rejection tests run on a CPU-only
+# machine -- no GPU required. Only the unlock test, which performs a real
+# read, needs cupy + CUDA.
+# ---------------------------------------------------------------------------
+
+
+def test_gpu_direct_fsspec_source_rejected_under_stable_only(memory_tiff_url):
+    """The direct ``read_geotiff_gpu`` entry point gates remote sources too."""
+    with pytest.raises(RemoteStableSourcesOnlyError) as excinfo:
+        read_geotiff_gpu(memory_tiff_url, stable_only=True)
+    _assert_remote_stable_error(excinfo)
+
+
+def test_gpu_direct_http_source_rejected_under_stable_only():
+    """An ``http://`` source is gated on the GPU path before any fetch."""
+    with pytest.raises(RemoteStableSourcesOnlyError) as excinfo:
+        read_geotiff_gpu("http://example.invalid/sample.tif", stable_only=True)
+    _assert_remote_stable_error(excinfo)
+
+
+@_gpu_only
+def test_gpu_direct_fsspec_source_allowed_with_experimental_optin(memory_tiff_url):
+    """``allow_experimental_codecs=True`` unlocks the advanced tier on GPU."""
+    result = read_geotiff_gpu(
+        memory_tiff_url, stable_only=True, allow_experimental_codecs=True,
+    )
+    assert result.shape == (4, 4)