diff --git a/ci/build_wheel.sh b/ci/build_wheel.sh
index 73d242fe..c3b35d82 100755
--- a/ci/build_wheel.sh
+++ b/ci/build_wheel.sh
@@ -14,6 +14,7 @@ package_type=$3
 #
 EXCLUDE_ARGS=(
     --exclude libcuda.so.1
+    --exclude "libnccl.so.*"
     --exclude libnvidia-ml.so.1
     --exclude librapids_logger.so
     --exclude librmm.so
diff --git a/ci/build_wheel_pylibwholegraph.sh b/ci/build_wheel_pylibwholegraph.sh
index 32b6901a..93b6d1cd 100755
--- a/ci/build_wheel_pylibwholegraph.sh
+++ b/ci/build_wheel_pylibwholegraph.sh
@@ -17,7 +17,7 @@ RAPIDS_PY_CUDA_SUFFIX="$(rapids-wheel-ctk-name-gen "${RAPIDS_CUDA_VERSION}")"
 LIBWHOLEGRAPH_WHEELHOUSE=$(RAPIDS_PY_WHEEL_NAME="libwholegraph_${RAPIDS_PY_CUDA_SUFFIX}" rapids-download-wheels-from-github cpp)
 echo "libwholegraph-${RAPIDS_PY_CUDA_SUFFIX} @ file://$(echo "${LIBWHOLEGRAPH_WHEELHOUSE}"/libwholegraph_*.whl)" >> "${PIP_CONSTRAINT}"
 
-export SKBUILD_CMAKE_ARGS="-DBUILD_SHARED_LIBS=ON;-DCMAKE_MESSAGE_LOG_LEVEL=VERBOSE;-DCUDA_STATIC_RUNTIME=ON;-DWHOLEGRAPH_BUILD_WHEELS=ON"
+export SKBUILD_CMAKE_ARGS="-DBUILD_SHARED_LIBS=ON;-DCMAKE_MESSAGE_LOG_LEVEL=VERBOSE;-DCUDA_STATIC_RUNTIME=ON"
 
 ./ci/build_wheel.sh pylibwholegraph ${package_dir} python
 ./ci/validate_wheel.sh ${package_dir} "${RAPIDS_WHEEL_BLD_OUTPUT_DIR}"
diff --git a/ci/test_wheel_pylibwholegraph.sh b/ci/test_wheel_pylibwholegraph.sh
index da40ea70..f59d200c 100755
--- a/ci/test_wheel_pylibwholegraph.sh
+++ b/ci/test_wheel_pylibwholegraph.sh
@@ -5,6 +5,12 @@ set -e          # abort the script on error
 set -o pipefail # piped commands propagate their error
 set -E          # ERR traps are inherited by subcommands
 
+# Delete system libnccl.so to ensure the wheel is used.
+# (but only do this in CI, to avoid breaking local dev environments)
+if [[ "${CI:-}" == "true" ]]; then
+  rm -rf /usr/lib64/libnccl*
+fi
+
 source rapids-init-pip
 
 RAPIDS_PY_CUDA_SUFFIX="$(rapids-wheel-ctk-name-gen ${RAPIDS_CUDA_VERSION})"
diff --git a/cpp/CMakeLists.txt b/cpp/CMakeLists.txt
index c82d2ef0..35bb1d7e 100644
--- a/cpp/CMakeLists.txt
+++ b/cpp/CMakeLists.txt
@@ -45,6 +45,7 @@ option(WHOLEGRAPH_EXCLUDE_NVSHMEM_FROM_ALL "Exclude nvshmem targets from wholeGr
        ON
 )
 option(BUILD_BENCHMARKS "Configure CMake to build benchmark" ON)
+option(USE_NCCL_RUNTIME_WHEEL "Use the NCCL wheel at runtime instead of the system library" OFF)
 
 # ##################################################################################################
 # * Set options based on user defined one  -----------------------------------
diff --git a/dependencies.yaml b/dependencies.yaml
index cba7125b..aba23208 100644
--- a/dependencies.yaml
+++ b/dependencies.yaml
@@ -21,6 +21,7 @@ files:
       - depends_on_dask_cudf
       - depends_on_cupy
       - depends_on_cuml
+      - depends_on_nccl
       - depends_on_pytorch
       - depends_on_pyg
       - python_run_cugraph_pyg
@@ -49,6 +50,7 @@ files:
       - test_cpp
       - depends_on_libwholegraph
       - depends_on_libwholegraph_tests
+      - depends_on_nccl
   test_notebooks:
     output: none
     includes:
@@ -105,8 +107,9 @@ files:
     extras:
       table: project
     includes:
-      - depends_on_rapids_logger
       - depends_on_libraft
+      - depends_on_nccl
+      - depends_on_rapids_logger
   py_build_pylibwholegraph:
     output: pyproject
     pyproject_dir: python/pylibwholegraph
@@ -342,7 +345,6 @@ dependencies:
       - output_types: conda
         packages:
           - *cmake_ver
-          - nccl>=2.19
   test_notebook:
     common:
       - output_types: [conda, requirements]
@@ -393,6 +395,21 @@ dependencies:
             packages:
               - pytorch>=2.3
           - {matrix: null, packages: ["pytorch>=2.3"]}
+  depends_on_nccl:
+    common:
+      - output_types: conda
+        packages:
+          - nccl>=2.19
+    specific:
+      - output_types: [pyproject, requirements]
+        matrices:
+          - matrix:
+              cuda: "12.*"
+              cuda_suffixed: "true"
+            packages:
+              - nvidia-nccl-cu12>=2.19
+          - matrix:
+            packages:
   depends_on_ogb:
     common:
       - output_types: [conda, requirements, pyproject]
diff --git a/python/cugraph-pyg/pyproject.toml b/python/cugraph-pyg/pyproject.toml
index 3bb02858..7a24d840 100644
--- a/python/cugraph-pyg/pyproject.toml
+++ b/python/cugraph-pyg/pyproject.toml
@@ -80,4 +80,4 @@ select = [
 ]
 
 # PyPI limit is 100 MiB, fail CI before we get too close to that
-max_allowed_size_compressed = '75M'
+max_allowed_size_compressed = '10Mi'
diff --git a/python/libwholegraph/CMakeLists.txt b/python/libwholegraph/CMakeLists.txt
index 20a7cbab..4ac98532 100644
--- a/python/libwholegraph/CMakeLists.txt
+++ b/python/libwholegraph/CMakeLists.txt
@@ -30,5 +30,16 @@ project(
 
 SET(BUILD_TESTS OFF)
 SET(BUILD_BENCHMARKS OFF)
+SET(USE_NCCL_RUNTIME_WHEEL ON)
 
 add_subdirectory(../../cpp/ libwholegraph)
+
+if(USE_NCCL_RUNTIME_WHEEL)
+  list(APPEND rpaths "$ORIGIN/../../nvidia/nccl/lib")
+endif()
+
+set_property(
+  TARGET wholegraph
+  PROPERTY INSTALL_RPATH ${rpaths}
+  APPEND
+)
diff --git a/python/libwholegraph/pyproject.toml b/python/libwholegraph/pyproject.toml
index e84cd2fe..bccbe59e 100644
--- a/python/libwholegraph/pyproject.toml
+++ b/python/libwholegraph/pyproject.toml
@@ -59,8 +59,8 @@ select = [
     "distro-too-large-compressed",
 ]
 
-# detect when package size grows significantly
-max_allowed_size_compressed = '0.4G'
+# PyPI limit is 100 MiB, fail CI before we get too close to that
+max_allowed_size_compressed = '80Mi'
 
 [tool.scikit-build]
 build-dir = "build/{wheel_tag}"
diff --git a/python/pylibwholegraph/pyproject.toml b/python/pylibwholegraph/pyproject.toml
index 0ff191e2..a036ba36 100644
--- a/python/pylibwholegraph/pyproject.toml
+++ b/python/pylibwholegraph/pyproject.toml
@@ -87,5 +87,5 @@ select = [
     "distro-too-large-compressed",
 ]
 
-# detect when package size grows significantly
-max_allowed_size_compressed = '400M'
+# PyPI limit is 100 MiB, fail CI before we get too close to that
+max_allowed_size_compressed = '10Mi'