diff --git a/osdc/clusters.yaml b/osdc/clusters.yaml
index 44aa9a75..f6e15a08 100644
--- a/osdc/clusters.yaml
+++ b/osdc/clusters.yaml
@@ -83,6 +83,10 @@ defaults:
     cpu_disruption_budget: "20%"
   buildkit:
     replicas_per_arch: 12
+    autoscaling:
+      enabled: false
+  keda:
+    chart_version: "2.16.1"
   monitoring:
     grafana_cloud_url: "https://prometheus-prod-36-prod-us-west-0.grafana.net/api/prom/push"
     grafana_cloud_read_url: "https://prometheus-prod-36-prod-us-west-0.grafana.net/api/prom"
@@ -162,8 +166,14 @@ clusters:
       github_secret_name: pytorch-arc-staging
       runner_name_prefix: "c-mt-"
     buildkit:
-      replicas_per_arch: 2
       arm64_instance_type: m7gd.16xlarge
+      arm64_pods_per_node: 4
+      autoscaling:
+        enabled: true
+        amd64_min: 2  # 1x m6id.24xlarge (2 pods/node)
+        amd64_max: 8
+        arm64_min: 4  # 1x m7gd.16xlarge (4 pods/node)
+        arm64_max: 8
     pypi_cache:
       replicas: 1
     modules:
@@ -171,6 +181,7 @@ clusters:
       - arc
       - nodepools
       - arc-runners
+      - keda
       - buildkit
       - pypi-cache
       - cache-enforcer
@@ -283,11 +294,15 @@ clusters:
       min_node_age_seconds: 900
     buildkit:
       amd64_instance_type: m6id.24xlarge
-      amd64_replicas: 32
       amd64_pods_per_node: 2
       arm64_instance_type: m7gd.16xlarge
-      arm64_replicas: 8
       arm64_pods_per_node: 4
+      autoscaling:
+        enabled: true
+        amd64_min: 2   # 1x m6id.24xlarge (2 pods/node)
+        amd64_max: 128  # 14d peak 105; headroom to spare above it
+        arm64_min: 4   # 1x m7gd.16xlarge (4 pods/node)
+        arm64_max: 16  # 14d peak 8, likely capped by fixed pool; headroom
     arc-runners:
       github_config_url: "https://github.com/pytorch"
       github_secret_name: pytorch-arc-cbr-production
diff --git a/osdc/integration-tests/docker/test-buildkit-scale/Dockerfile b/osdc/integration-tests/docker/test-buildkit-scale/Dockerfile
new file mode 100644
index 00000000..e317eb61
--- /dev/null
+++ b/osdc/integration-tests/docker/test-buildkit-scale/Dockerfile
@@ -0,0 +1,6 @@
+# Holds one BuildKit slot (~10 min) so a burst of parallel builds exceeds the
+# warm baseline and forces KEDA scale-up. CACHEBUST differs per build so each
+# actually runs (no layer reuse).
+FROM alpine:3.21
+ARG CACHEBUST
+RUN echo "osdc buildkit scale-test ${CACHEBUST}" && sleep 600
diff --git a/osdc/integration-tests/scripts/python/phases.py b/osdc/integration-tests/scripts/python/phases.py
index 27ddebf6..17654000 100644
--- a/osdc/integration-tests/scripts/python/phases.py
+++ b/osdc/integration-tests/scripts/python/phases.py
@@ -276,15 +276,17 @@ def prepare_pr(
     # Write integration test workflow
     (workflows_dir / "integration-test.yaml").write_text(workflow_content)
 
-    # Copy build-image reusable workflow
-    build_wf_src = upstream_dir / "integration-tests" / "workflows" / "build-image.yaml"
-    (workflows_dir / "build-image.yaml").write_text(build_wf_src.read_text())
-
-    # Copy test Dockerfile
-    docker_dir = canary_path / "docker" / "test-buildkit"
-    docker_dir.mkdir(parents=True, exist_ok=True)
-    dockerfile_src = upstream_dir / "integration-tests" / "docker" / "test-buildkit" / "Dockerfile"
-    (docker_dir / "Dockerfile").write_text(dockerfile_src.read_text())
+    # Copy reusable BuildKit workflows
+    wf_root = upstream_dir / "integration-tests" / "workflows"
+    for wf in ("build-image.yaml", "build-image-scale.yaml"):
+        (workflows_dir / wf).write_text((wf_root / wf).read_text())
+
+    # Copy test Dockerfiles (connectivity + autoscaling scale test)
+    docker_root = upstream_dir / "integration-tests" / "docker"
+    for name in ("test-buildkit", "test-buildkit-scale"):
+        dst = canary_path / "docker" / name
+        dst.mkdir(parents=True, exist_ok=True)
+        (dst / "Dockerfile").write_text((docker_root / name / "Dockerfile").read_text())
 
     # Commit
     run_cmd(["git", "add", "-A"], cwd=canary_path)
diff --git a/osdc/integration-tests/scripts/python/test_run.py b/osdc/integration-tests/scripts/python/test_run.py
index ff3a7bcb..dc4d5188 100644
--- a/osdc/integration-tests/scripts/python/test_run.py
+++ b/osdc/integration-tests/scripts/python/test_run.py
@@ -112,11 +112,13 @@ def workflow_template(tmp_path):
     )
     (wf_dir / "integration-test.yaml.tpl").write_text(template)
 
-    # Also create build-image.yaml and Dockerfile for prepare_pr
+    # Also create reusable workflows and Dockerfiles for prepare_pr
     (wf_dir / "build-image.yaml").write_text("name: build-image\n")
-    docker_dir = upstream / "integration-tests" / "docker" / "test-buildkit"
-    docker_dir.mkdir(parents=True)
-    (docker_dir / "Dockerfile").write_text("FROM alpine\n")
+    (wf_dir / "build-image-scale.yaml").write_text("name: build-image-scale\n")
+    for name in ("test-buildkit", "test-buildkit-scale"):
+        docker_dir = upstream / "integration-tests" / "docker" / name
+        docker_dir.mkdir(parents=True)
+        (docker_dir / "Dockerfile").write_text("FROM alpine\n")
 
     return upstream
 
diff --git a/osdc/integration-tests/workflows/build-image-scale.yaml b/osdc/integration-tests/workflows/build-image-scale.yaml
new file mode 100644
index 00000000..b485dc5a
--- /dev/null
+++ b/osdc/integration-tests/workflows/build-image-scale.yaml
@@ -0,0 +1,59 @@
+# Reusable workflow: BuildKit autoscaling scale test.
+# Launches a burst of 8 parallel builds against one arch's remote BuildKit, each
+# holding a slot (server maxconn=1) for ~10 min via a sleep. The warm baseline
+# (amd64_min / arm64_min) is below the burst, so the builds finish within
+# timeout-minutes only if KEDA scales the pool up. Without scale-up they
+# serialize through the baseline pods and the back of the queue times out — i.e.
+# this job FAILS when autoscaling does not happen.
+name: BuildKit Scale Test
+
+on:
+  workflow_call:
+    inputs:
+      arch:
+        description: "Target architecture (amd64 or arm64)"
+        required: true
+        type: string
+      runner_label:
+        description: "Runner label to use (includes cluster prefix)"
+        required: false
+        type: string
+        default: "l-x86iavx512-2-4"
+
+jobs:
+  scale:
+    # Runs on x86 — BuildKit is a *remote* builder; arch selects the endpoint.
+    # timeout-minutes is the gate: scaled-up ~18 min, serialized ~43 min.
+    runs-on: ${{ inputs.runner_label }}
+    timeout-minutes: 30
+    strategy:
+      fail-fast: false
+      matrix:
+        replica: [1, 2, 3, 4, 5, 6, 7, 8]
+    container:
+      image: ghcr.io/actions/actions-runner:latest
+    steps:
+      - name: Install buildctl
+        run: |
+          BUILDKIT_VERSION="v0.29.0"
+          mkdir -p "$HOME/.local/bin"
+          curl -sSL "https://github.com/moby/buildkit/releases/download/${BUILDKIT_VERSION}/buildkit-${BUILDKIT_VERSION}.linux-amd64.tar.gz" \
+            | tar xz --strip-components=1 -C "$HOME/.local/bin" bin/buildctl
+          echo "$HOME/.local/bin" >> "$GITHUB_PATH"
+          "$HOME/.local/bin/buildctl" --version
+
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Occupy a BuildKit slot (~10 min) to drive autoscaling
+        run: |
+          ENDPOINT="tcp://buildkitd-${{ inputs.arch }}.buildkit:1234"
+          echo "Build ${{ matrix.replica }} -> $ENDPOINT"
+          buildctl --addr "$ENDPOINT" build \
+            --frontend dockerfile.v0 \
+            --local context=docker/test-buildkit-scale \
+            --local dockerfile=docker/test-buildkit-scale \
+            --opt build-arg:CACHEBUST=${{ inputs.arch }}-${{ matrix.replica }}-${{ github.run_id }} \
+            --no-cache \
+            --output type=cacheonly
+          echo "PASS: build ${{ matrix.replica }} finished within timeout"
diff --git a/osdc/integration-tests/workflows/integration-test.yaml.tpl b/osdc/integration-tests/workflows/integration-test.yaml.tpl
index c9563480..7d4fb3aa 100644
--- a/osdc/integration-tests/workflows/integration-test.yaml.tpl
+++ b/osdc/integration-tests/workflows/integration-test.yaml.tpl
@@ -1514,6 +1514,20 @@ jobs:
       arch: arm64
       runner_label: {{PREFIX}}l-x86iamx-8-32
 
+  # ── BuildKit Autoscaling Scale Test ───────────────────────────────────
+  # Bursts 8 parallel builds per arch; fails if KEDA does not scale the pool up.
+  buildkit-scale-amd64:
+    uses: ./.github/workflows/build-image-scale.yaml
+    with:
+      arch: amd64
+      runner_label: {{PREFIX}}l-x86iamx-8-32
+
+  buildkit-scale-arm64:
+    uses: ./.github/workflows/build-image-scale.yaml
+    with:
+      arch: arm64
+      runner_label: {{PREFIX}}l-x86iamx-8-32
+
   # ── Harbor Cache Test ─────────────────────────────────────────────────
   test-harbor:
     runs-on: {{PREFIX}}l-x86iamx-8-32
diff --git a/osdc/modules/buildkit/README.md b/osdc/modules/buildkit/README.md
new file mode 100644
index 00000000..21fd44dd
--- /dev/null
+++ b/osdc/modules/buildkit/README.md
@@ -0,0 +1,40 @@
+# BuildKit module
+
+Remote BuildKit build service: per-arch `buildkitd` Deployments behind an HAProxy
+LB, on dedicated Karpenter NodePools. Clients build with
+`buildctl --addr tcp://buildkitd-<arch>.buildkit:1234`.
+
+Sizing is per-arch in `clusters.yaml` (`buildkit.{amd64,arm64}_{replicas,pods_per_node}`,
+`*_instance_type`); pod CPU/mem is computed by `scripts/python/generate_buildkit.py`.
+
+## Autoscaling (optional, `buildkit.autoscaling.enabled`)
+
+Absorb bursts of concurrent builds without overloading existing pods, and scale
+back to a small warm baseline when idle.
+
+- **One build per pod** — HAProxy `server maxconn 1` (matches buildkitd
+  `max-parallelism = 1`). Excess builds **queue** in HAProxy instead of stacking
+  on a busy pod; as new pods register (DNS), queued builds flow onto them, so
+  scaled-up pods never sit idle. `timeout queue` must stay set and large enough
+  to outlast a node-provision cycle — if omitted it falls back to `timeout
+  connect` (5s), which would abort queued builds before pods scale up.
+- **In-cluster scale signal** — KEDA `ScaledObject` per arch, `metrics-api`
+  scraping the LB's own metrics (`haproxy_backend_current_sessions`) — no external
+  metrics backend.
+- **Warm baseline** — `amd64_min` / `arm64_min` keep ≥1 node per arch up so the
+  common case gets a free warm pod immediately. `*_max` caps the burst; NodePool
+  limits are sized to `*_max`.
+- **Safe scale-down** — `preStop` drain (waits until the pod's `:1234` is idle)
+  + long `terminationGracePeriodSeconds` + PDB, so a build is never killed
+  mid-flight.
+- **Node consolidation** — the NodePool uses Karpenter `consolidationPolicy:
+  WhenEmpty`: a node is reclaimed only once it has no buildkitd pod, never by
+  evicting a running build to bin-pack. So after a burst, scattered survivor
+  pods can leave nodes half-full (more nodes than the cold baseline) until they
+  drain naturally — a deliberate trade of some idle node cost for zero build
+  disruption.
+
+Build clients should retry the connect so a build can wait for a pod from a cold
+or queued pool.
+
+Requires the `keda` module deployed before `buildkit` (provides the CRDs).
diff --git a/osdc/modules/buildkit/deploy.sh b/osdc/modules/buildkit/deploy.sh
index c7be6d39..d7c1fa72 100755
--- a/osdc/modules/buildkit/deploy.sh
+++ b/osdc/modules/buildkit/deploy.sh
@@ -34,22 +34,38 @@ AMD64_REPLICAS=$(uv run "$CFG" "$CLUSTER" buildkit.amd64_replicas "$REPLICAS")
 ARM64_REPLICAS=$(uv run "$CFG" "$CLUSTER" buildkit.arm64_replicas "$REPLICAS")
 AMD64_PODS_PER_NODE=$(uv run "$CFG" "$CLUSTER" buildkit.amd64_pods_per_node "$PODS_PER_NODE")
 ARM64_PODS_PER_NODE=$(uv run "$CFG" "$CLUSTER" buildkit.arm64_pods_per_node "$PODS_PER_NODE")
+AUTOSCALING=$(uv run "$CFG" "$CLUSTER" buildkit.autoscaling.enabled false | tr '[:upper:]' '[:lower:]')
 
 GENERATED_DIR="$MODULE_DIR/generated"
 
 # --- Generate manifests ---
 
 echo "Generating BuildKit manifests..."
-uv run "$MODULE_DIR/scripts/python/generate_buildkit.py" \
-  --arm64-instance-type "$ARM64_INSTANCE" \
-  --amd64-instance-type "$AMD64_INSTANCE" \
-  --replicas "$REPLICAS" \
-  --pods-per-node "$PODS_PER_NODE" \
-  --amd64-replicas "$AMD64_REPLICAS" \
-  --arm64-replicas "$ARM64_REPLICAS" \
-  --amd64-pods-per-node "$AMD64_PODS_PER_NODE" \
-  --arm64-pods-per-node "$ARM64_PODS_PER_NODE" \
+GEN_ARGS=(
+  --arm64-instance-type "$ARM64_INSTANCE"
+  --amd64-instance-type "$AMD64_INSTANCE"
+  --replicas "$REPLICAS"
+  --pods-per-node "$PODS_PER_NODE"
+  --amd64-replicas "$AMD64_REPLICAS"
+  --arm64-replicas "$ARM64_REPLICAS"
+  --amd64-pods-per-node "$AMD64_PODS_PER_NODE"
+  --arm64-pods-per-node "$ARM64_PODS_PER_NODE"
   --output-dir "$GENERATED_DIR"
+)
+if [[ "$AUTOSCALING" == "true" ]]; then
+  AMD64_MIN=$(uv run "$CFG" "$CLUSTER" buildkit.autoscaling.amd64_min 2)
+  AMD64_MAX=$(uv run "$CFG" "$CLUSTER" buildkit.autoscaling.amd64_max 8)
+  ARM64_MIN=$(uv run "$CFG" "$CLUSTER" buildkit.autoscaling.arm64_min 4)
+  ARM64_MAX=$(uv run "$CFG" "$CLUSTER" buildkit.autoscaling.arm64_max 8)
+  GEN_ARGS+=(
+    --autoscaling
+    --amd64-min "$AMD64_MIN"
+    --amd64-max "$AMD64_MAX"
+    --arm64-min "$ARM64_MIN"
+    --arm64-max "$ARM64_MAX"
+  )
+fi
+uv run "$MODULE_DIR/scripts/python/generate_buildkit.py" "${GEN_ARGS[@]}"
 
 # --- Apply NodePools (with cluster name substitution) ---
 
@@ -93,5 +109,13 @@ else
   kubectl rollout status deployment/buildkitd-amd64 -n buildkit --timeout=15m
 fi
 
+# --- KEDA autoscaling (optional) ---
+# Scales on the in-cluster buildkit LB metrics; no external metrics backend.
+
+if [[ "$AUTOSCALING" == "true" ]]; then
+  echo "Applying KEDA autoscaling manifests..."
+  kubectl_apply_if_changed -f "$GENERATED_DIR/autoscaling.yaml"
+fi
+
 echo "BuildKit deployed."
 kubectl get pods -n buildkit -o wide
diff --git a/osdc/modules/buildkit/kubernetes/base/drain-configmap.yaml b/osdc/modules/buildkit/kubernetes/base/drain-configmap.yaml
new file mode 100644
index 00000000..89395ae9
--- /dev/null
+++ b/osdc/modules/buildkit/kubernetes/base/drain-configmap.yaml
@@ -0,0 +1,21 @@
+apiVersion: v1
+kind: ConfigMap
+metadata:
+  name: buildkitd-drain
+  namespace: buildkit
+data:
+  # preStop drain: block termination until no in-flight build remains. A build
+  # keeps an ESTABLISHED inbound connection on :1234 for its whole duration;
+  # require two consecutive idle polls so a transient health check can't be
+  # mistaken for "done". terminationGracePeriodSeconds caps the total wait.
+  drain.sh: |
+    #!/bin/sh
+    idle=0
+    while [ "$idle" -lt 2 ]; do
+      if netstat -tn 2>/dev/null | awk '$NF=="ESTABLISHED" && $4 ~ /:1234$/{f=1} END{exit !f}'; then
+        idle=0
+      else
+        idle=$((idle + 1))
+      fi
+      sleep 15
+    done
diff --git a/osdc/modules/buildkit/kubernetes/base/haproxy.yaml b/osdc/modules/buildkit/kubernetes/base/haproxy.yaml
index 3bfd4eb2..cd0b25ee 100644
--- a/osdc/modules/buildkit/kubernetes/base/haproxy.yaml
+++ b/osdc/modules/buildkit/kubernetes/base/haproxy.yaml
@@ -24,6 +24,9 @@ data:
       timeout connect 5s
       timeout client 120m
       timeout server 120m
+      # Queue builds (server maxconn=1) while pods scale up. Keep <= 120m, the max
+      # time a docker build is allowed to run.
+      timeout queue 60m
       log global
       option tcplog
 
@@ -71,15 +74,15 @@ data:
     # resolution warnings.
     backend bk_arm64
       balance leastconn
-      server-template arm 10 buildkitd-arm64-pods.buildkit.svc.cluster.local:1234 check resolvers k8s init-addr none resolve-prefer ipv6
+      server-template arm 10 buildkitd-arm64-pods.buildkit.svc.cluster.local:1234 check resolvers k8s init-addr none resolve-prefer ipv6 maxconn 1
 
     backend bk_amd64
       balance leastconn
-      server-template amd 10 buildkitd-amd64-pods.buildkit.svc.cluster.local:1234 check resolvers k8s init-addr none resolve-prefer ipv6
+      server-template amd 10 buildkitd-amd64-pods.buildkit.svc.cluster.local:1234 check resolvers k8s init-addr none resolve-prefer ipv6 maxconn 1
 
     backend bk_all
       balance leastconn
-      server-template all 20 buildkitd-pods.buildkit.svc.cluster.local:1234 check resolvers k8s init-addr none resolve-prefer ipv6
+      server-template all 20 buildkitd-pods.buildkit.svc.cluster.local:1234 check resolvers k8s init-addr none resolve-prefer ipv6 maxconn 1
 
 ---
 apiVersion: apps/v1
diff --git a/osdc/modules/buildkit/kubernetes/base/kustomization.yaml b/osdc/modules/buildkit/kubernetes/base/kustomization.yaml
index 9cbffb38..32572408 100644
--- a/osdc/modules/buildkit/kubernetes/base/kustomization.yaml
+++ b/osdc/modules/buildkit/kubernetes/base/kustomization.yaml
@@ -5,6 +5,8 @@ kind: Kustomization
 resources:
   - namespace.yaml
   - configmap.yaml
+  - drain-configmap.yaml
   - haproxy.yaml
   - service.yaml
   - networkpolicy.yaml
+  - poddisruptionbudget.yaml
diff --git a/osdc/modules/buildkit/kubernetes/base/poddisruptionbudget.yaml b/osdc/modules/buildkit/kubernetes/base/poddisruptionbudget.yaml
new file mode 100644
index 00000000..b717bb64
--- /dev/null
+++ b/osdc/modules/buildkit/kubernetes/base/poddisruptionbudget.yaml
@@ -0,0 +1,25 @@
+# Cap voluntary disruptions (node consolidation, drains) to one builder per arch
+# at a time so evictions go through the preStop drain instead of killing builds.
+apiVersion: policy/v1
+kind: PodDisruptionBudget
+metadata:
+  name: buildkitd-arm64
+  namespace: buildkit
+spec:
+  maxUnavailable: 1
+  selector:
+    matchLabels:
+      app: buildkitd
+      arch: arm64
+---
+apiVersion: policy/v1
+kind: PodDisruptionBudget
+metadata:
+  name: buildkitd-amd64
+  namespace: buildkit
+spec:
+  maxUnavailable: 1
+  selector:
+    matchLabels:
+      app: buildkitd
+      arch: amd64
diff --git a/osdc/modules/buildkit/scripts/python/generate_buildkit.py b/osdc/modules/buildkit/scripts/python/generate_buildkit.py
index e7468418..788dd9e6 100644
--- a/osdc/modules/buildkit/scripts/python/generate_buildkit.py
+++ b/osdc/modules/buildkit/scripts/python/generate_buildkit.py
@@ -103,6 +103,7 @@ def generate_deployment_yaml(
     arm64_replicas: int | None = None,
     amd64_pods_per_node: int | None = None,
     arm64_pods_per_node: int | None = None,
+    autoscaling: bool = False,
 ) -> str:
     """Generate the combined Deployment YAML for both architectures.
 
@@ -118,6 +119,35 @@ def generate_deployment_yaml(
     arm64_res = compute_pod_resources(arm64_instance, arm64_pods_per_node)
     amd64_res = compute_pod_resources(amd64_instance, amd64_pods_per_node)
 
+    # Under autoscaling KEDA owns replicas: omit `replicas`, add a preStop drain.
+    grace_line = "      terminationGracePeriodSeconds: 8100\n" if autoscaling else ""
+    lifecycle_block = (
+        """
+          lifecycle:
+            preStop:
+              exec:
+                command: ["/bin/sh", "/opt/drain/drain.sh"]"""
+        if autoscaling
+        else ""
+    )
+    drain_mount = (
+        """
+            - name: drain
+              mountPath: /opt/drain
+              readOnly: true"""
+        if autoscaling
+        else ""
+    )
+    drain_volume = (
+        """
+        - name: drain
+          configMap:
+            name: buildkitd-drain
+            defaultMode: 0555"""
+        if autoscaling
+        else ""
+    )
+
     log_info(
         f"arm64 ({arm64_instance}): {arm64_res['cpu']} vCPU, {arm64_res['memory_gi']}Gi per pod "
         f"(allocatable: {arm64_res['allocatable_cpu_m']}m CPU, {arm64_res['allocatable_mem_mi']}Mi mem)"
@@ -128,6 +158,7 @@ def generate_deployment_yaml(
     )
 
     def _deployment_block(arch, instance_type, cpu, memory_gi, replicas, pods_per_node):
+        replicas_line = "" if autoscaling else f"  replicas: {replicas}\n"
         return f"""apiVersion: apps/v1
 kind: Deployment
 metadata:
@@ -139,8 +170,7 @@ def _deployment_block(arch, instance_type, cpu, memory_gi, replicas, pods_per_no
     app.kubernetes.io/name: buildkitd
     app.kubernetes.io/component: build-service
 spec:
-  replicas: {replicas}
-  strategy:
+{replicas_line}  strategy:
     type: RollingUpdate
     rollingUpdate:
       maxSurge: 0
@@ -155,7 +185,7 @@ def _deployment_block(arch, instance_type, cpu, memory_gi, replicas, pods_per_no
         app: buildkitd
         arch: {arch}
     spec:
-      nodeSelector:
+{grace_line}      nodeSelector:
         workload-type: buildkit
         instance-type: "{instance_type}"
 
@@ -208,7 +238,7 @@ def _deployment_block(arch, instance_type, cpu, memory_gi, replicas, pods_per_no
                   fieldPath: metadata.name
 
           securityContext:
-            privileged: true
+            privileged: true{lifecycle_block}
 
           readinessProbe:
             exec:
@@ -237,7 +267,7 @@ def _deployment_block(arch, instance_type, cpu, memory_gi, replicas, pods_per_no
               subPathExpr: $(POD_NAME)
             - name: git-cache
               mountPath: /opt/git-cache
-              readOnly: true
+              readOnly: true{drain_mount}
 
       volumes:
         - name: config
@@ -252,7 +282,7 @@ def _deployment_block(arch, instance_type, cpu, memory_gi, replicas, pods_per_no
         - name: git-cache
           hostPath:
             path: /mnt/k8s-disks/0/git-cache
-            type: DirectoryOrCreate"""
+            type: DirectoryOrCreate{drain_volume}"""
 
     arm64_block = _deployment_block(
         "arm64", arm64_instance, arm64_res["cpu"], arm64_res["memory_gi"], arm64_replicas, arm64_pods_per_node
@@ -453,6 +483,63 @@ def _nodepool_block(arch, instance_type, cpu_limit, memory_limit_gi):
     return arm64_block + "\n\n---\n" + amd64_block + "\n"
 
 
+def generate_autoscaling_yaml(
+    amd64_min: int,
+    amd64_max: int,
+    arm64_min: int,
+    arm64_max: int,
+) -> str:
+    """Generate per-arch KEDA ScaledObjects.
+
+    Each arch scales on its HAProxy backend's active build count
+    (haproxy_backend_current_sessions), scraped in-cluster from the buildkit LB
+    metrics endpoint — no external metrics backend. With server maxconn=1, the LB
+    queues bursts while KEDA/Karpenter bring up pods; minReplicaCount keeps a warm
+    baseline so the common case has a free pod immediately.
+    """
+
+    metrics_url = "http://buildkitd-lb-metrics.buildkit.svc.cluster.local:9404/metrics"
+
+    def _scaledobject(arch, backend, min_replicas, max_replicas):
+        return f"""apiVersion: keda.sh/v1alpha1
+kind: ScaledObject
+metadata:
+  name: buildkitd-{arch}
+  namespace: buildkit
+spec:
+  scaleTargetRef:
+    name: buildkitd-{arch}
+  minReplicaCount: {min_replicas}
+  maxReplicaCount: {max_replicas}
+  cooldownPeriod: 600
+  advanced:
+    horizontalPodAutoscalerConfig:
+      behavior:
+        scaleDown:
+          stabilizationWindowSeconds: 600
+          policies:
+            - type: Pods
+              value: 1
+              periodSeconds: 120
+  triggers:
+    - type: metrics-api
+      metadata:
+        url: "{metrics_url}"
+        format: "prometheus"
+        valueLocation: 'haproxy_backend_current_sessions{{proxy="{backend}"}}'
+        targetValue: "1"
+"""
+
+    header = "# KEDA autoscaling — auto-generated by generate_buildkit.py. Do not edit by hand.\n"
+    return (
+        header
+        + "\n"
+        + _scaledobject("amd64", "bk_amd64", amd64_min, amd64_max)
+        + "\n---\n"
+        + _scaledobject("arm64", "bk_arm64", arm64_min, arm64_max)
+    )
+
+
 def main():
     parser = argparse.ArgumentParser(description="Generate BuildKit Deployment and NodePool YAMLs")
     parser.add_argument("--arm64-instance-type", required=True, help="ARM64 instance type (e.g., m8gd.24xlarge)")
@@ -464,8 +551,17 @@ def main():
     parser.add_argument("--amd64-pods-per-node", type=int, default=None, help="Override amd64 pods per node")
     parser.add_argument("--arm64-pods-per-node", type=int, default=None, help="Override arm64 pods per node")
     parser.add_argument("--output-dir", required=True, help="Output directory for generated YAMLs")
+    parser.add_argument("--autoscaling", action="store_true", help="Generate KEDA autoscaling manifests")
+    parser.add_argument("--amd64-min", type=int, default=0, help="KEDA minReplicaCount for amd64")
+    parser.add_argument("--amd64-max", type=int, default=0, help="KEDA maxReplicaCount for amd64")
+    parser.add_argument("--arm64-min", type=int, default=0, help="KEDA minReplicaCount for arm64")
+    parser.add_argument("--arm64-max", type=int, default=0, help="KEDA maxReplicaCount for arm64")
     args = parser.parse_args()
 
+    if args.autoscaling and not (args.amd64_max and args.arm64_max):
+        log_error("--autoscaling requires --amd64-max and --arm64-max")
+        return 1
+
     # Validate instance types
     for it in [args.arm64_instance_type, args.amd64_instance_type]:
         if it not in INSTANCE_SPECS:
@@ -498,26 +594,51 @@ def main():
         arm64_replicas=args.arm64_replicas,
         amd64_pods_per_node=args.amd64_pods_per_node,
         arm64_pods_per_node=args.arm64_pods_per_node,
+        autoscaling=args.autoscaling,
     )
     deployment_path = output_dir / "deployment.yaml"
     deployment_path.write_text(deployment_yaml)
     log_info(f"Wrote {deployment_path}")
 
-    # Generate nodepools
-    nodepools_yaml = generate_nodepools_yaml(
-        args.arm64_instance_type,
-        args.amd64_instance_type,
-        args.replicas,
-        args.pods_per_node,
-        amd64_replicas=args.amd64_replicas,
-        arm64_replicas=args.arm64_replicas,
-        amd64_pods_per_node=args.amd64_pods_per_node,
-        arm64_pods_per_node=args.arm64_pods_per_node,
-    )
+    # Size NodePool limits for the peak: the per-arch autoscaling ceiling
+    # (amd64_max / arm64_max) when enabled, otherwise the per-arch replica counts.
+    if args.autoscaling:
+        nodepools_yaml = generate_nodepools_yaml(
+            args.arm64_instance_type,
+            args.amd64_instance_type,
+            args.replicas,
+            args.pods_per_node,
+            amd64_replicas=args.amd64_max,
+            arm64_replicas=args.arm64_max,
+            amd64_pods_per_node=args.amd64_pods_per_node,
+            arm64_pods_per_node=args.arm64_pods_per_node,
+        )
+    else:
+        nodepools_yaml = generate_nodepools_yaml(
+            args.arm64_instance_type,
+            args.amd64_instance_type,
+            args.replicas,
+            args.pods_per_node,
+            amd64_replicas=args.amd64_replicas,
+            arm64_replicas=args.arm64_replicas,
+            amd64_pods_per_node=args.amd64_pods_per_node,
+            arm64_pods_per_node=args.arm64_pods_per_node,
+        )
     nodepools_path = output_dir / "nodepools.yaml"
     nodepools_path.write_text(nodepools_yaml)
     log_info(f"Wrote {nodepools_path}")
 
+    if args.autoscaling:
+        autoscaling_yaml = generate_autoscaling_yaml(
+            args.amd64_min,
+            args.amd64_max,
+            args.arm64_min,
+            args.arm64_max,
+        )
+        autoscaling_path = output_dir / "autoscaling.yaml"
+        autoscaling_path.write_text(autoscaling_yaml)
+        log_info(f"Wrote {autoscaling_path}")
+
     return 0
 
 
diff --git a/osdc/modules/buildkit/scripts/python/test_generate_buildkit.py b/osdc/modules/buildkit/scripts/python/test_generate_buildkit.py
index 387a283d..c373c9c0 100644
--- a/osdc/modules/buildkit/scripts/python/test_generate_buildkit.py
+++ b/osdc/modules/buildkit/scripts/python/test_generate_buildkit.py
@@ -12,6 +12,7 @@
     DAEMONSET_OVERHEAD_MEM_MI,
     MARGIN,
     compute_pod_resources,
+    generate_autoscaling_yaml,
     generate_deployment_yaml,
     generate_nodepools_yaml,
 )
@@ -341,6 +342,65 @@ def test_nodepool_limits_scale_per_arch(self):
         assert np["buildkit-arm64"]["spec"]["limits"]["cpu"] == "256"
 
 
+# ============================================================================
+# autoscaling
+# ============================================================================
+
+
+class TestAutoscalingDeployment:
+    """When autoscaling=True the Deployment yields control of replicas to KEDA
+    and gains a preStop drain so scale-down never kills an in-flight build."""
+
+    def test_omits_replicas(self):
+        output = generate_deployment_yaml("m8gd.24xlarge", "m6id.24xlarge", 4, 2, autoscaling=True)
+        for d in parse_all_yaml(output):
+            assert "replicas" not in d["spec"]
+
+    def test_keeps_replicas_when_disabled(self):
+        output = generate_deployment_yaml("m8gd.24xlarge", "m6id.24xlarge", 4, 2)
+        for d in parse_all_yaml(output):
+            assert d["spec"]["replicas"] == 4
+
+    def test_drain_prestop_and_grace(self):
+        output = generate_deployment_yaml("m8gd.24xlarge", "m6id.24xlarge", 4, 2, autoscaling=True)
+        for d in parse_all_yaml(output):
+            spec = d["spec"]["template"]["spec"]
+            assert spec["terminationGracePeriodSeconds"] == 8100
+            container = spec["containers"][0]
+            assert container["lifecycle"]["preStop"]["exec"]["command"] == ["/bin/sh", "/opt/drain/drain.sh"]
+            assert "drain" in {vm["name"] for vm in container["volumeMounts"]}
+            assert "drain" in {v["name"] for v in spec["volumes"]}
+
+
+class TestGenerateAutoscalingYaml:
+    """Tests for generate_autoscaling_yaml — in-cluster KEDA ScaledObjects."""
+
+    def _docs(self):
+        output = generate_autoscaling_yaml(2, 8, 4, 8)
+        return parse_all_yaml(output)
+
+    def test_only_scaledobjects_no_external_auth(self):
+        # In-cluster signal → no TriggerAuthentication / Grafana secret needed.
+        kinds = sorted(d["kind"] for d in self._docs())
+        assert kinds == ["ScaledObject", "ScaledObject"]
+
+    def test_per_arch_min_max(self):
+        scaled = {d["metadata"]["name"]: d for d in self._docs() if d["kind"] == "ScaledObject"}
+        assert set(scaled) == {"buildkitd-amd64", "buildkitd-arm64"}
+        assert scaled["buildkitd-amd64"]["spec"]["minReplicaCount"] == 2
+        assert scaled["buildkitd-amd64"]["spec"]["maxReplicaCount"] == 8
+        assert scaled["buildkitd-arm64"]["spec"]["minReplicaCount"] == 4
+        assert scaled["buildkitd-arm64"]["spec"]["maxReplicaCount"] == 8
+
+    def test_in_cluster_metrics_api_trigger(self):
+        scaled = {d["metadata"]["name"]: d for d in self._docs() if d["kind"] == "ScaledObject"}
+        for name, backend in [("buildkitd-amd64", "bk_amd64"), ("buildkitd-arm64", "bk_arm64")]:
+            trig = scaled[name]["spec"]["triggers"][0]
+            assert trig["type"] == "metrics-api"
+            assert "buildkitd-lb-metrics.buildkit" in trig["metadata"]["url"]
+            assert f'proxy="{backend}"' in trig["metadata"]["valueLocation"]
+
+
 # ============================================================================
 # generate_nodepools_yaml
 # ============================================================================
@@ -520,6 +580,64 @@ def test_deployment_yaml_parseable(self, tmp_path):
         docs = parse_all_yaml(deployment_text)
         assert len(docs) == 2
 
+    def test_autoscaling_writes_manifests(self, tmp_path):
+        output_dir = tmp_path / "output"
+
+        import generate_buildkit
+
+        test_args = [
+            "generate_buildkit.py",
+            "--arm64-instance-type",
+            "m8gd.24xlarge",
+            "--amd64-instance-type",
+            "m6id.24xlarge",
+            "--replicas",
+            "1",
+            "--pods-per-node",
+            "2",
+            "--output-dir",
+            str(output_dir),
+            "--autoscaling",
+            "--amd64-min",
+            "2",
+            "--amd64-max",
+            "8",
+            "--arm64-min",
+            "4",
+            "--arm64-max",
+            "8",
+        ]
+        with patch.object(sys, "argv", test_args):
+            result = generate_buildkit.main()
+
+        assert result == 0
+        autoscaling = parse_all_yaml((output_dir / "autoscaling.yaml").read_text())
+        assert sorted(d["kind"] for d in autoscaling) == ["ScaledObject", "ScaledObject"]
+
+    def test_autoscaling_requires_params(self, tmp_path):
+        output_dir = tmp_path / "output"
+
+        import generate_buildkit
+
+        test_args = [
+            "generate_buildkit.py",
+            "--arm64-instance-type",
+            "m8gd.24xlarge",
+            "--amd64-instance-type",
+            "m6id.24xlarge",
+            "--replicas",
+            "1",
+            "--pods-per-node",
+            "2",
+            "--output-dir",
+            str(output_dir),
+            "--autoscaling",
+        ]
+        with patch.object(sys, "argv", test_args):
+            result = generate_buildkit.main()
+
+        assert result == 1
+
     def test_unknown_instance_type_fails(self, tmp_path):
         output_dir = tmp_path / "output"
         test_args = [
diff --git a/osdc/modules/keda/deploy.sh b/osdc/modules/keda/deploy.sh
new file mode 100755
index 00000000..af16c6ab
--- /dev/null
+++ b/osdc/modules/keda/deploy.sh
@@ -0,0 +1,35 @@
+#!/usr/bin/env bash
+set -euo pipefail
+#
+# KEDA module deploy script.
+# Called by: just deploy-module <cluster> keda
+# Args: $1=cluster-id  $2=cluster-name  $3=region
+#
+# Installs the KEDA operator (provides the ScaledObject/TriggerAuthentication
+# CRDs that the buildkit module uses to autoscale builders).
+
+CLUSTER="$1"
+MODULE_DIR="$(cd "$(dirname "$0")" && pwd)"
+REPO_ROOT="${OSDC_ROOT:-$(cd "$MODULE_DIR/../.." && pwd)}"
+UPSTREAM_ROOT="${OSDC_UPSTREAM:-$REPO_ROOT}"
+# shellcheck source=/dev/null
+source "$UPSTREAM_ROOT/scripts/mise-activate.sh"
+# shellcheck source=/dev/null
+source "$UPSTREAM_ROOT/scripts/helm-upgrade.sh"
+CFG="$UPSTREAM_ROOT/scripts/cluster-config.py"
+
+NAMESPACE="keda"
+CHART_VERSION=$(uv run "$CFG" "$CLUSTER" keda.chart_version 2.16.1)
+
+helm repo add kedacore https://kedacore.github.io/charts >/dev/null 2>&1 || true
+helm repo update kedacore >/dev/null 2>&1 || true
+
+helm_upgrade_if_changed keda "$NAMESPACE" \
+  --create-namespace \
+  --version "$CHART_VERSION" \
+  -f "$MODULE_DIR/helm/values.yaml" \
+  --timeout 10m \
+  --wait \
+  kedacore/keda
+
+echo "KEDA deployed (chart $CHART_VERSION)."
diff --git a/osdc/modules/keda/helm/values.yaml b/osdc/modules/keda/helm/values.yaml
new file mode 100644
index 00000000..2d151e9f
--- /dev/null
+++ b/osdc/modules/keda/helm/values.yaml
@@ -0,0 +1,4 @@
+# Schedule on base-infra nodes; all other nodes are reserved by taints.
+tolerations:
+  - key: CriticalAddonsOnly
+    operator: Exists