diff --git a/alembic/versions/20260927_0047_tag_vertical_counts.py b/alembic/versions/20260927_0047_tag_vertical_counts.py
new file mode 100644
index 0000000..0c32f6b
--- /dev/null
+++ b/alembic/versions/20260927_0047_tag_vertical_counts.py
@@ -0,0 +1,41 @@
+"""Liczniki tagow per wertykal (JAV, animacja), liczone w tle zamiast na kazde wejscie.
+
+/tags?vertical=jav liczyl na zywo: sceny z zywym zrodlem JAV, ich scene_tags, GROUP BY,
+i to dwa razy (licznik total i strona). Na rozgrzanym cache 1 s, na zimnym 17-24 s
+(log API 2026-09-27, zakladka JAV w apce). Tabela (vertical, tag_id) zamiast kolumny
+na `tags` na kazdy wertykal, zeby nowy wertykal nie wymagal migracji. Wypelnia ja
+refresh_taxonomy_counts (co 3 h); dopoki dla wertykalu nie ma wierszy, API liczy po staremu.
+
+Revision ID: 0047_tag_vertical_counts
+Revises: 0046_scenes_title_stats
+Create Date: 2026-09-27
+"""
+from __future__ import annotations
+
+import sqlalchemy as sa
+from alembic import op
+from sqlalchemy.dialects.postgresql import UUID
+
+revision = "0047_tag_vertical_counts"
+down_revision: str | None = "0046_scenes_title_stats"
+branch_labels = None
+depends_on = None
+
+
+def upgrade() -> None:
+    op.create_table(
+        "tag_vertical_counts",
+        sa.Column("vertical", sa.String(32), nullable=False),
+        sa.Column("tag_id", UUID(as_uuid=True), nullable=False),
+        sa.Column("scene_count", sa.Integer, nullable=False),
+        sa.PrimaryKeyConstraint("vertical", "tag_id", name="pk_tag_vertical_counts"),
+        sa.ForeignKeyConstraint(
+            ["tag_id"], ["tags.id"],
+            name="fk_tag_vertical_counts_tag_id_tags",
+            ondelete="CASCADE",
+        ),
+    )
+
+
+def downgrade() -> None:
+    op.drop_table("tag_vertical_counts")
diff --git a/app/api/taxonomies.py b/app/api/taxonomies.py
index c444d74..78ff2ab 100644
--- a/app/api/taxonomies.py
+++ b/app/api/taxonomies.py
@@ -27,6 +27,7 @@ from app.models.playback_source import PlaybackSource
 from app.models.scene import ScenePerformer, SceneTag
 from app.models.studio import Studio
 from app.models.tag import Tag
+from app.models.tag_vertical_count import TagVerticalCount
 from app.verticals import VERTICALS
 
 log = logging.getLogger(__name__)
@@ -158,25 +159,36 @@ def list_tags(
         return TagListOut(items=items, total=total, page=page, per_page=per_page)
 
     if vertical in VERTICALS:
-        # Wertykał liczymy NA ŻYWO, bez denormalizacji: to małe zbiory (animacja ~9,5 tys.
-        # scen wobec 3,8 mln całości), więc agregat jest tani, a dokładanie kolumny
-        # `scene_count_<wertykał>` do `tags` kosztowałoby migrację przy każdym nowym
-        # wertykale. Zdenormalizowany `Tag.scene_count` opisuje WYŁĄCZNIE główny feed.
-        v_scenes = (
-            select(PlaybackSource.scene_id)
-            .where(
-                PlaybackSource.dead_at.is_(None),
-                PlaybackSource.origin.in_(VERTICALS[vertical]),
+        # Liczniki wertykału czytamy z tag_vertical_counts (refresh_taxonomy_counts co 3 h,
+        # migracja 0047). Agregat na żywo trwał 1 s na rozgrzanym cache i 17-24 s na
+        # zimnym (zakładka JAV, 2026-09-27). Zostaje jako zapas, gdy tabela jest jeszcze
+        # pusta dla tego wertykału, np. zaraz po wdrożeniu.
+        # Zdenormalizowany `Tag.scene_count` opisuje WYŁĄCZNIE główny feed.
+        has_counts = session.execute(
+            select(TagVerticalCount.tag_id).where(TagVerticalCount.vertical == vertical).limit(1)
+        ).first()
+        if has_counts:
+            v_count = (
+                select(TagVerticalCount.tag_id, TagVerticalCount.scene_count.label("c"))
+                .where(TagVerticalCount.vertical == vertical)
+                .subquery()
+            )
+        else:
+            v_scenes = (
+                select(PlaybackSource.scene_id)
+                .where(
+                    PlaybackSource.dead_at.is_(None),
+                    PlaybackSource.origin.in_(VERTICALS[vertical]),
+                )
+                .distinct()
+                .subquery()
+            )
+            v_count = (
+                select(SceneTag.tag_id, func.count().label("c"))
+                .join(v_scenes, v_scenes.c.scene_id == SceneTag.scene_id)
+                .group_by(SceneTag.tag_id)
+                .subquery()
             )
-            .distinct()
-            .subquery()
-        )
-        v_count = (
-            select(SceneTag.tag_id, func.count().label("c"))
-            .join(v_scenes, v_scenes.c.scene_id == SceneTag.scene_id)
-            .group_by(SceneTag.tag_id)
-            .subquery()
-        )
         base = (
             select(Tag, func.coalesce(v_count.c.c, 0).label("scene_count"))
             .outerjoin(v_count, v_count.c.tag_id == Tag.id)
diff --git a/app/models/__init__.py b/app/models/__init__.py
index 89a7c8d..9ba6619 100644
--- a/app/models/__init__.py
+++ b/app/models/__init__.py
@@ -38,6 +38,7 @@ from app.models.scene import (
 from app.models.source import Source
 from app.models.studio import Studio, StudioAlias, StudioExternalRef
 from app.models.tag import Tag
+from app.models.tag_vertical_count import TagVerticalCount
 
 __all__ = [
     "Base",
@@ -84,4 +85,5 @@ __all__ = [
     "StudioAlias",
     "StudioExternalRef",
     "Tag",
+    "TagVerticalCount",
 ]
diff --git a/app/models/tag_vertical_count.py b/app/models/tag_vertical_count.py
new file mode 100644
index 0000000..a3e19f5
--- /dev/null
+++ b/app/models/tag_vertical_count.py
@@ -0,0 +1,25 @@
+"""Liczba scen tagu w wertykale (JAV, animacja), odswiezana w tle.
+
+Wypelnia `refresh_taxonomy_counts`, czyta /tags?vertical=. Ta sama definicja co dawny
+agregat na zywo: sceny z >= 1 zywym zrodlem z originow wertykalu, bez progu dlugosci.
+Patrz migracja 0047.
+"""
+from __future__ import annotations
+
+import uuid
+
+from sqlalchemy import ForeignKey, Integer, String
+from sqlalchemy.dialects.postgresql import UUID
+from sqlalchemy.orm import Mapped, mapped_column
+
+from app.models.base import Base
+
+
+class TagVerticalCount(Base):
+    __tablename__ = "tag_vertical_counts"
+
+    vertical: Mapped[str] = mapped_column(String(32), primary_key=True)
+    tag_id: Mapped[uuid.UUID] = mapped_column(
+        UUID(as_uuid=True), ForeignKey("tags.id", ondelete="CASCADE"), primary_key=True
+    )
+    scene_count: Mapped[int] = mapped_column(Integer, nullable=False)
diff --git a/app/scheduler/taxonomy_counts.py b/app/scheduler/taxonomy_counts.py
index 071e9b2..dcc7069 100644
--- a/app/scheduler/taxonomy_counts.py
+++ b/app/scheduler/taxonomy_counts.py
@@ -23,7 +23,7 @@ import logging
 from sqlalchemy import text
 
 from app.db import session_scope
-from app.verticals import ALL_VERTICAL_ORIGINS
+from app.verticals import ALL_VERTICAL_ORIGINS, VERTICALS
 
 log = logging.getLogger(__name__)
 
@@ -129,11 +129,35 @@ _STUDIOS_SQL = text(
 )
 
 
+# Tagi wertykalu (lista w zakladce JAV/animacja): sceny z zywym zrodlem z jego originow,
+# bez progu dlugosci, ta sama definicja co dawny agregat w /tags?vertical=. Podmiana
+# calego kompletu w jednej transakcji, to kilka tysiecy wierszy (migracja 0047).
+_VERTICAL_TAGS_DELETE_SQL = text("DELETE FROM tag_vertical_counts WHERE vertical = :vertical")
+_VERTICAL_TAGS_INSERT_SQL = text(
+    """
+    INSERT INTO tag_vertical_counts (vertical, tag_id, scene_count)
+    SELECT :vertical, st.tag_id, count(*)
+    FROM scene_tags st
+    JOIN (
+        SELECT DISTINCT ps.scene_id FROM playback_sources ps
+        WHERE ps.dead_at IS NULL AND ps.origin = ANY(:origins)
+    ) v ON v.scene_id = st.scene_id
+    GROUP BY st.tag_id
+    """
+)
+
+
 def refresh_taxonomy_counts() -> dict[str, int]:
     """Przelicza scene_count dla tags/performers/studios. Zwraca rowcount per tabela
     (ile wierszy faktycznie się zmieniło)."""
     out: dict[str, int] = {}
     with session_scope() as session:
+        for vertical, origins in VERTICALS.items():
+            session.execute(_VERTICAL_TAGS_DELETE_SQL, {"vertical": vertical})
+            out[f"tags_{vertical}"] = session.execute(
+                _VERTICAL_TAGS_INSERT_SQL, {"vertical": vertical, "origins": list(origins)}
+            ).rowcount or 0
+            session.commit()
         for name, stmt in (
             ("tags", _TAGS_SQL),
             ("performers", _PERFORMERS_SQL),