diff --git a/python/sglang/srt/layers/moe/utils.py b/python/sglang/srt/layers/moe/utils.py
index 78bd3abd5..c6ac46ee8 100644
--- a/python/sglang/srt/layers/moe/utils.py
+++ b/python/sglang/srt/layers/moe/utils.py
@@ -279,6 +279,11 @@ def initialize_moe_config(server_args: ServerArgs):
     DEEPEP_CONFIG = server_args.deepep_config or ""
     IS_TBO_ENABLED = server_args.enable_two_batch_overlap
     IS_SBO_ENABLED = server_args.enable_single_batch_overlap
+    if IS_SBO_ENABLED and torch.cuda.is_available():
+        if torch.cuda.get_device_capability()[0] == 9:
+            raise ValueError(
+                "SBO (single batch overlap) is not supported on SM90 GPUs with latest sgl-deep-gemm wheel. Please try removing --enable-single-batch-overlap argument."
+            )
     TBO_TOKEN_DISTRIBUTION_THRESHOLD = server_args.tbo_token_distribution_threshold
     DISABLE_FLASHINFER_CUTLASS_MOE_FP4_ALLGATHER = (
         server_args.disable_flashinfer_cutlass_moe_fp4_allgather