diff --git a/services/capture/src/capture-manager.js b/services/capture/src/capture-manager.js
index 7db5a6e..45e4bf2 100644
--- a/services/capture/src/capture-manager.js
+++ b/services/capture/src/capture-manager.js
@@ -181,8 +181,24 @@ function parseFps(framerate, fallback = 60) {
   return Number.isFinite(f) && f > 0 ? f : fallback;
 }
 
+// Which physical GPU this sidecar's NVENC encodes should use. node-agent
+// round-robins capture ports across the host's GPUs and passes the index here.
+// We MUST select it explicitly with ffmpeg's `-gpu N` because the capture
+// sidecars run Privileged (so they see every /dev/nvidiaN regardless of
+// NVIDIA_VISIBLE_DEVICES) — without -gpu, nvenc defaults every session to GPU 0
+// and all 8 ports pile onto one card → it falls below realtime → video freezes.
+const CAPTURE_GPU_INDEX = (() => {
+  const v = process.env.CAPTURE_GPU_INDEX;
+  if (v == null || v === '' || v === 'all') return null;
+  const n = parseInt(v, 10);
+  return Number.isInteger(n) && n >= 0 ? n : null;
+})();
+// `-gpu N` must come BEFORE the input/encoder is initialized; ffmpeg accepts it
+// as an encoder option right after -c:v. Returns [] when no pin is configured.
+const nvencGpuSel = () => (CAPTURE_GPU_INDEX != null ? ['-gpu', String(CAPTURE_GPU_INDEX)] : []);
+
 function hevcNvencArgs(framerate, growing) {
-  const base = ['-c:v', 'hevc_nvenc', '-preset', 'p4', '-rc', 'vbr', '-profile:v', 'main10'];
+  const base = ['-c:v', 'hevc_nvenc', ...nvencGpuSel(), '-preset', 'p4', '-rc', 'vbr', '-profile:v', 'main10'];
   if (growing) {
     return [...base, '-bf', '0', '-forced-idr', '1', '-g', '600', '-force_key_frames', 'expr:1'];
   }
@@ -232,7 +248,7 @@ function buildHlsVideoArgs(masterCodec, framerate) {
     // Low-latency NVENC preset (p1 + ll tune). forced-idr + a keyframe every GOP
     // frames keeps segment boundaries on IDR frames so hls.js can sync cleanly.
     return [
-      '-c:v', 'h264_nvenc', '-preset', 'p1', '-tune', 'll',
+      '-c:v', 'h264_nvenc', ...nvencGpuSel(), '-preset', 'p1', '-tune', 'll',
       '-pix_fmt', 'yuv420p', '-b:v', '2M',
       '-g', String(gop), '-forced-idr', '1', '-sc_threshold', '0',
     ];
diff --git a/services/node-agent/index.js b/services/node-agent/index.js
index 40ff530..51ff3ad 100644
--- a/services/node-agent/index.js
+++ b/services/node-agent/index.js
@@ -527,6 +527,10 @@ async function handleSidecarStart(body, res) {
       var startVisibleDevices = pickVisibleDevices(gpuUuid, capturePort);
       sidecarEnv.push(`NVIDIA_VISIBLE_DEVICES=${startVisibleDevices}`);
       sidecarEnv.push('NVIDIA_DRIVER_CAPABILITIES=video,compute,utility');
+      // Privileged sidecars see every /dev/nvidiaN regardless of
+      // NVIDIA_VISIBLE_DEVICES, so also tell ffmpeg explicitly which GPU to
+      // encode on via CAPTURE_GPU_INDEX (capture-manager adds `-gpu N`).
+      if (startVisibleDevices !== 'all') sidecarEnv.push(`CAPTURE_GPU_INDEX=${startVisibleDevices}`);
       console.log(`[gpu] sidecar port ${capturePort} → NVIDIA_VISIBLE_DEVICES=${startVisibleDevices}`);
     }
 
@@ -841,6 +845,7 @@ async function handleSidecarStandby(body, res) {
       standbyVisibleDevices = pickVisibleDevices(gpuUuid, capturePort);
       sidecarEnv.push(`NVIDIA_VISIBLE_DEVICES=${standbyVisibleDevices}`);
       sidecarEnv.push('NVIDIA_DRIVER_CAPABILITIES=video,compute,utility');
+      if (standbyVisibleDevices !== 'all') sidecarEnv.push(`CAPTURE_GPU_INDEX=${standbyVisibleDevices}`);
       console.log(`[gpu] standby sidecar port ${capturePort} → NVIDIA_VISIBLE_DEVICES=${standbyVisibleDevices}`);
     }
     sidecarEnv.push(`FC_URL=${FC_URL}`);