diff --git a/vllm/nvfp4_cutedsl.py b/vllm/nvfp4_cutedsl.py
index 3d73cad5..e39cf8d4 100644
--- a/vllm/nvfp4_cutedsl.py
+++ b/vllm/nvfp4_cutedsl.py
@@ -138,10 +138,9 @@ class CuTeDSLMoERunner:
             self.max_num_tokens * self.top_k, device=self.device
         )
         
-        # Padded hidden/activated: per-layer, sized for max capture budget
-        # NOT sized for max_num_tokens (8192) which would be too much for 60 layers
-        max_slots = self.max_num_tokens * self.top_k
-        padded_max_slots = ((max_slots + 127) // 128) * 128
+        # Padded hidden/activated: sized for num_experts * max_chunks * 128 (the actual padded size)
+        max_rows_per_expert = self._max_chunks_per_expert * 128
+        padded_max_slots = self.num_experts * max_rows_per_expert
         self._padded_hidden_buf = torch.zeros(
             padded_max_slots, self.hidden_size, dtype=torch.bfloat16, device=self.device
         )