diff --git a/src/server/offloading_algo/offloading_algo.py b/src/server/offloading_algo/offloading_algo.py
index c62bb38..d2ebcfc 100644
--- a/src/server/offloading_algo/offloading_algo.py
+++ b/src/server/offloading_algo/offloading_algo.py
@@ -1,3 +1,4 @@
+import itertools
 import json
 import logging
 from pathlib import Path
@@ -149,8 +150,6 @@ def mixed_computation_evaluation(self):
 
         # ── OPTIMIZATION: Use prefix sums to avoid recomputing sum() per layer ──
         # This reduces the inner loop from O(N²) to O(N).
-        import itertools
-
         device_prefix = list(itertools.accumulate(self.inference_time_device))
         edge_prefix = list(itertools.accumulate(self.inference_time_edge))
         total_edge = edge_prefix[self.num_layers - 1] if self.num_layers > 0 else 0
@@ -190,7 +189,10 @@ def device_only_evaluation(self):
         if _DEBUG_ENABLED:
             logger.debug("Performing Device Only Offloading")
         initial_cost = sum(self.inference_time_device[: self.num_layers])
-        layer_data_size = self.layers_sizes[self.num_layers - 1]
+        # A fully local inference never transmits intermediate layer data
+        # over the network, so no network cost applies here (unlike the
+        # split strategies, which do send data to the edge).
+        layer_data_size = 0
         edge_computation_cost = 0
 
         # No Offloading: Device Only Computation
@@ -204,7 +206,7 @@ def device_only_evaluation(self):
             strategy="device_only",
             layer=self.num_layers - 1,
             device_compute_cost=initial_cost,
-            network_cost=layer_data_size / (self.avg_speed or 1),
+            network_cost=0,
             edge_compute_cost=edge_computation_cost,
             estimated_total_cost=last_evaluation,
             considered_for_selection=True,
diff --git a/tests/test_offloading_algo/test_offloading_algo.py b/tests/test_offloading_algo/test_offloading_algo.py
index 0d1d7a4..3139c31 100644
--- a/tests/test_offloading_algo/test_offloading_algo.py
+++ b/tests/test_offloading_algo/test_offloading_algo.py
@@ -41,5 +41,28 @@ def test_offloading_fixture_falls_back_when_sample_file_is_missing():
     assert _load_sample_values("test_samples/missing.json", [1.0, 2.0]) == [1.0, 2.0]
 
 
+def test_device_only_evaluation_has_no_network_cost():
+    # A fully local (device-only) inference never transmits intermediate
+    # layer data over the network, so it must not incur a network cost.
+    offloading_algo = OffloadingAlgo(
+        avg_speed=10.0,
+        num_layers=3,
+        layers_sizes=[100, 200, 300],
+        inference_time_device=[1.0, 1.0, 1.0],
+        inference_time_edge=[1.0, 1.0, 1.0],
+    )
+    offloading_algo.device_only_evaluation()
+
+    device_only_candidate = next(
+        c
+        for c in offloading_algo.candidate_evaluations
+        if c["strategy"] == "device_only"
+    )
+    assert device_only_candidate["network_cost"] == 0
+    assert device_only_candidate["estimated_total_cost"] == sum(
+        [1.0, 1.0, 1.0]
+    )
+
+
 if __name__ == "__main__":
     pytest.main()