diff --git a/scripts/dispatcher.py b/scripts/dispatcher.py
index 6f9f516..4e4e730 100644
--- a/scripts/dispatcher.py
+++ b/scripts/dispatcher.py
@@ -103,25 +103,37 @@ def worker_free_vram_mib(worker: dict) -> int:
         return 0
 
 
+_jobs_per_worker: dict[str, int] = {}
+
+
 def pick_worker(estimated_vram_mib: int) -> dict | None:
-    """Pick worker with most effective free VRAM (actual free minus local reservations)
-    and reserve the estimated VRAM. Returns None if none fit."""
+    """Pick a worker that (a) has enough free VRAM AND (b) is currently the least busy —
+    otherwise the beefiest GPU always wins and the second worker sits idle."""
     with _worker_lock:
-        best = None
+        candidates = []
         for w in WORKERS:
             free = worker_free_vram_mib(w) - _reserved_vram.get(w["host"], 0)
-            if free >= estimated_vram_mib and (best is None or free > best[0]):
-                best = (free, w)
-        if best:
-            _reserved_vram[best[1]["host"]] = _reserved_vram.get(best[1]["host"], 0) + estimated_vram_mib
-            return best[1]
-    return None
+            print(f"  pick_worker: {w['host']} free={free} reserved={_reserved_vram.get(w['host'], 0)} load={_jobs_per_worker.get(w['host'], 0)}", flush=True)
+            if free < estimated_vram_mib:
+                continue
+            load = _jobs_per_worker.get(w["host"], 0)
+            # Sort key must avoid comparing dicts: use host as tiebreaker.
+            candidates.append(((load, -free, w["host"]), w))
+        if not candidates:
+            print(f"  pick_worker: no candidate for {estimated_vram_mib} MiB", flush=True)
+            return None
+        candidates.sort(key=lambda c: c[0])
+        w = candidates[0][1]
+        _reserved_vram[w["host"]] = _reserved_vram.get(w["host"], 0) + estimated_vram_mib
+        _jobs_per_worker[w["host"]] = _jobs_per_worker.get(w["host"], 0) + 1
+        return w
 
 
 def release_worker(worker: dict, estimated_vram_mib: int):
     with _worker_lock:
         h = worker["host"]
         _reserved_vram[h] = max(0, _reserved_vram.get(h, 0) - estimated_vram_mib)
+        _jobs_per_worker[h] = max(0, _jobs_per_worker.get(h, 0) - 1)
 
 
 def estimate_vram_mib(frame_count: int) -> int: