diff --git a/examples/cuda-fanout/gpu-worker.lsp b/examples/cuda-fanout/gpu-worker.lsp
index b22e571..d65c53a 100644
--- a/examples/cuda-fanout/gpu-worker.lsp
+++ b/examples/cuda-fanout/gpu-worker.lsp
@@ -701,7 +701,11 @@
 ;; Running max per-cell VRAM observed. Conservative — uses largest
 ;; bin seen, not average. Single-slot mutable. Initial 4096 MiB =
 ;; pessimistic seed until we observe real cells.
-(define *vram-per-cell-max-mib* 4096)
+;; Seed lowered 4096→2500 (2026-06-11) — real K=2 bin avg is ~2.4GB.
+;; Overprojection at 4096 throttled admit to 2-3 concurrent for big bins
+;; even when 4-5 would fit; lower seed lets autoscaler use full GPU headroom.
+;; record-cell-vram! in child still updates running max if larger seen.
+(define *vram-per-cell-max-mib* 2500)
 
 (define (vram-used-mib)
   ;; nvidia-smi --query-gpu=memory.used --format=csv,noheader,nounits