diff --git a/cpp/src/arrow/gpu/cuda_context.h b/cpp/src/arrow/gpu/cuda_context.h
index 00bcf94c8b6ab..0115ed19a103d 100644
--- a/cpp/src/arrow/gpu/cuda_context.h
+++ b/cpp/src/arrow/gpu/cuda_context.h
@@ -76,7 +76,7 @@ class ARROW_EXPORT CudaDeviceManager {
   static std::unique_ptr<CudaDeviceManager> instance_;
 
   class Impl;
-  std::unique_ptr<Impl> impl_;
+  std::shared_ptr<Impl> impl_;
 
   friend class CudaContext;
   friend class CudaDevice;
@@ -146,7 +146,7 @@ class ARROW_EXPORT CudaDevice : public Device {
   /// \endcond
 
   explicit CudaDevice(Impl);
-  std::unique_ptr<Impl> impl_;
+  std::shared_ptr<Impl> impl_;
 };
 
 /// \brief Return whether a device instance is a CudaDevice
@@ -297,7 +297,7 @@ class ARROW_EXPORT CudaContext : public std::enable_shared_from_this<CudaContext
                                    uintptr_t dst, uintptr_t src, int64_t nbytes);
 
   class Impl;
-  std::unique_ptr<Impl> impl_;
+  std::shared_ptr<Impl> impl_;
 
   friend class CudaBuffer;
   friend class CudaBufferReader;
diff --git a/cpp/src/arrow/gpu/cuda_internal.h b/cpp/src/arrow/gpu/cuda_internal.h
index 25eb6e06cee40..d70873634f869 100644
--- a/cpp/src/arrow/gpu/cuda_internal.h
+++ b/cpp/src/arrow/gpu/cuda_internal.h
@@ -33,6 +33,7 @@ namespace internal {
 
 std::string CudaErrorDescription(CUresult err);
 
+ARROW_EXPORT
 Status StatusFromCuda(CUresult res, const char* function_name = nullptr);
 
 #define CU_RETURN_NOT_OK(FUNC_NAME, STMT)                               \
diff --git a/cpp/src/arrow/gpu/cuda_memory.h b/cpp/src/arrow/gpu/cuda_memory.h
index 4efd3889407e1..18c23a507805a 100644
--- a/cpp/src/arrow/gpu/cuda_memory.h
+++ b/cpp/src/arrow/gpu/cuda_memory.h
@@ -250,10 +250,12 @@ Result<std::shared_ptr<CudaHostBuffer>> AllocateCudaHostBuffer(int device_number
                                                                const int64_t size);
 
 /// Low-level: get a device address through which the CPU data be accessed.
+ARROW_EXPORT
 Result<uintptr_t> GetDeviceAddress(const uint8_t* cpu_data,
                                    const std::shared_ptr<CudaContext>& ctx);
 
 /// Low-level: get a CPU address through which the device data be accessed.
+ARROW_EXPORT
 Result<uint8_t*> GetHostAddress(uintptr_t device_ptr);
 
 }  // namespace cuda