diff --git a/modules/nvidia_plugin/src/cuda_graph_context.cpp b/modules/nvidia_plugin/src/cuda_graph_context.cpp
index 4033b5df5b..a4374be63e 100644
--- a/modules/nvidia_plugin/src/cuda_graph_context.cpp
+++ b/modules/nvidia_plugin/src/cuda_graph_context.cpp
@@ -49,22 +49,8 @@ void CudaGraphInfo::set_graph(const CUDA::Graph& graph) {
     graphExec_.emplace(graph);
 }
 
-void CudaGraphInfo::set_params_graph(const CUDA::Graph& graph) {
-    paramsGraph_.emplace(graph);
-    paramsGraphExec_.emplace(graph);
-}
-
-void CudaGraphInfo::set_results_graph(const CUDA::Graph& graph) {
-    resultsGraph_.emplace(graph);
-    resultsGraphExec_.emplace(graph);
-}
-
 void CudaGraphInfo::launch(const CUDA::Stream& stream) const { graphExec_.value().launch(stream); }
 
-void CudaGraphInfo::launch_params_graph(const CUDA::Stream& stream) const { paramsGraphExec_.value().launch(stream); }
-
-void CudaGraphInfo::launch_results_graph(const CUDA::Stream& stream) const { resultsGraphExec_.value().launch(stream); }
-
 bool operator==(const CudaGraphInfo& lhs, const CudaGraphInfo& rhs) {
     return lhs.graph_ == rhs.graph_ && lhs.graphExec_ == rhs.graphExec_ && lhs.parameterNodes_ == rhs.parameterNodes_ &&
            lhs.resultNodes_ == rhs.resultNodes_ && lhs.transferNodes_ == rhs.transferNodes_ &&
diff --git a/modules/nvidia_plugin/src/cuda_graph_context.hpp b/modules/nvidia_plugin/src/cuda_graph_context.hpp
index 484bc15c58..295a8fbadd 100644
--- a/modules/nvidia_plugin/src/cuda_graph_context.hpp
+++ b/modules/nvidia_plugin/src/cuda_graph_context.hpp
@@ -51,12 +51,8 @@ class CudaGraphInfo {
     std::size_t get_kernels_count() const { return kernelNodes_.size(); }
 
     void set_graph(const CUDA::Graph& graph);
-    void set_params_graph(const CUDA::Graph& graph);
-    void set_results_graph(const CUDA::Graph& graph);
 
     void launch(const CUDA::Stream& stream) const;
-    void launch_params_graph(const CUDA::Stream& stream) const;
-    void launch_results_graph(const CUDA::Stream& stream) const;
 
     friend bool operator==(const CudaGraphInfo& lhs, const CudaGraphInfo& rhs);
     friend bool operator!=(const CudaGraphInfo& lhs, const CudaGraphInfo& rhs);
@@ -65,12 +61,6 @@ class CudaGraphInfo {
     std::optional<CUDA::Graph> graph_{};
     std::optional<CUDA::GraphExec> graphExec_{};
 
-    std::optional<CUDA::Graph> paramsGraph_{};
-    std::optional<CUDA::GraphExec> paramsGraphExec_{};
-
-    std::optional<CUDA::Graph> resultsGraph_{};
-    std::optional<CUDA::GraphExec> resultsGraphExec_{};
-
     std::map<std::string, CUDA::UploadNode> parameterNodes_;
     std::map<std::string, CUDA::DownloadNode> resultNodes_;
 
diff --git a/modules/nvidia_plugin/src/ops/tensor_iterator.cpp b/modules/nvidia_plugin/src/ops/tensor_iterator.cpp
index 4eeb991724..0d5eee11bb 100644
--- a/modules/nvidia_plugin/src/ops/tensor_iterator.cpp
+++ b/modules/nvidia_plugin/src/ops/tensor_iterator.cpp
@@ -226,23 +226,9 @@ void TensorIteratorOp::Capture(InferenceRequestContext& context,
     auto& mutableBuffer = workbuffers.mutable_buffers.at(0);
     auto& executionDelegator = context.getExecutionDelegator();
     executionDelegator.set_stream(stream);
-
     auto& graphInfo = context.getCudaGraphContext().get_current_graph_info();
+
     CUDA::GraphCapture capture{stream};
-    {
-        auto scope = capture.getScope();
-        // First iteration
-        for (const auto inputIdx : invariant_inputs_) {
-            const auto paramIdx = inputs_parameters_map_.at(inputIdx);
-            transferParam(stream, mutableBuffer, inputTensors, 0, inputIdx, paramIdx);
-        }
-        for (const auto& [inputIdx, paramIdx] : inputs_parameters_map_) {
-            if (portmap_inputs_.count(inputIdx) == 0) {
-                transferParam(stream, mutableBuffer, inputTensors, 0, inputIdx, paramIdx);
-            }
-        }
-    }
-    graphInfo.set_params_graph(capture.getGraph());
     {
         auto scope = capture.getScope();
         // Input mapping of ports
@@ -264,17 +250,6 @@ void TensorIteratorOp::Capture(InferenceRequestContext& context,
         }
     }
     graphInfo.set_graph(capture.getGraph());
-    {
-        auto scope = capture.getScope();
-        // Copy data to output
-        if (iterations_results_map_.count(num_iterations_ - 1) > 0) {
-            for (const auto& resultIdx : iterations_results_map_.at(num_iterations_ - 1)) {
-                const auto& outputIdx = results_outputs_map_.at(resultIdx);
-                transferResult(stream, mutableBuffer, outputTensors, num_iterations_ - 1, resultIdx, outputIdx);
-            }
-        }
-    }
-    graphInfo.set_results_graph(capture.getGraph());
 }
 
 void TensorIteratorOp::ExecuteGraph(InferenceRequestContext& context,
@@ -285,13 +260,22 @@ void TensorIteratorOp::ExecuteGraph(InferenceRequestContext& context,
     const auto& memoryManager = *memory_manager_;
     const auto& mutableBuffer = workbuffers.mutable_buffers.at(0);
 
-    auto& graphInfo = context.getCudaGraphContext().get_current_graph_info();
-
-    graphInfo.launch_params_graph(stream);
+    // First iteration; this part doesn't use CUDA graphs yet
+    for (const auto inputIdx : invariant_inputs_) {
+        const auto paramIdx = inputs_parameters_map_.at(inputIdx);
+        transferParam(stream, mutableBuffer, inputTensors, 0, inputIdx, paramIdx);
+    }
+    for (const auto& [inputIdx, paramIdx] : inputs_parameters_map_) {
+        if (portmap_inputs_.count(inputIdx) == 0) {
+            transferParam(stream, mutableBuffer, inputTensors, 0, inputIdx, paramIdx);
+        }
+    }
 
+    auto& graphInfo = context.getCudaGraphContext().get_current_graph_info();
     OPENVINO_ASSERT(graphInfo.get_kernels_count() == slices_.size() + inserts_.size(),
                     "CudaGraphContext/TensorIteratorOp slices or inserts count incosistency");
 
+    // TI body loop
     for (int64_t iter = 0; iter < num_iterations_; ++iter) {
         for (std::size_t i = 0; i < slices_.size(); ++i) {
             slices_[i].update_kernel_node(graphInfo, i, mutableBuffer, inputTensors, iter);
@@ -302,7 +286,13 @@ void TensorIteratorOp::ExecuteGraph(InferenceRequestContext& context,
         graphInfo.launch(stream);
     }
 
-    graphInfo.launch_results_graph(stream);
+    // Copy data to output; this part doesn't use CUDA graphs yet
+    if (iterations_results_map_.count(num_iterations_ - 1) > 0) {
+        for (const auto& resultIdx : iterations_results_map_.at(num_iterations_ - 1)) {
+            const auto& outputIdx = results_outputs_map_.at(resultIdx);
+            transferResult(stream, mutableBuffer, outputTensors, num_iterations_ - 1, resultIdx, outputIdx);
+        }
+    }
 }
 
 TensorIteratorOp::SliceLauncher::SliceLauncher(const TensorIteratorOp& ti, uint64_t inputIdx, uint64_t paramIdx)