diff --git a/services/node-agent/index.js b/services/node-agent/index.js
index 150cfec..26e4205 100644
--- a/services/node-agent/index.js
+++ b/services/node-agent/index.js
@@ -2,7 +2,7 @@ import http          from 'http';
 import os            from 'os';
 import fs            from 'fs';
 import crypto        from 'crypto';
-import { spawn, execFileSync } from 'child_process';
+import { spawn }     from 'child_process';
 
 const MAM_API_URL  = (process.env.MAM_API_URL  || 'http://localhost:3000').replace(/\/$/, '');
 const NODE_TOKEN   =  process.env.NODE_TOKEN   || '';
@@ -32,20 +32,16 @@ const VERSION      = '1.4.0';
 // 8 ports = 16 sessions hammering one card → it falls below realtime → the
 // framecache ring laps → video freezes/stutters then recovers. Pinning each
 // sidecar to GPU (port % CAPTURE_GPU_COUNT) spreads the load across all cards.
-// Auto-detected from nvidia-smi at boot; override with CAPTURE_GPU_COUNT.
-let CAPTURE_GPU_COUNT = parseInt(process.env.CAPTURE_GPU_COUNT || '0', 10) || 0;
+// GPU count comes from _gpuCache (populated at startup via probeGpusViaSmi,
+// which runs nvidia-smi in a throwaway GPU container because the node-agent
+// image itself has no nvidia-smi). CAPTURE_GPU_COUNT env overrides everything.
 function detectGpuCount() {
-  if (CAPTURE_GPU_COUNT > 0) return CAPTURE_GPU_COUNT;
-  try {
-    const out = execFileSync('nvidia-smi', ['--query-gpu=index', '--format=csv,noheader'], { timeout: 5000 })
-      .toString().trim();
-    const n = out ? out.split('\n').filter(Boolean).length : 0;
-    CAPTURE_GPU_COUNT = n > 0 ? n : 1;
-  } catch (_) {
-    CAPTURE_GPU_COUNT = 1; // no nvidia-smi / no GPU — single-device fallback
-  }
-  console.log(`[gpu] capture encodes will spread across ${CAPTURE_GPU_COUNT} GPU(s)`);
-  return CAPTURE_GPU_COUNT;
+  const envN = parseInt(process.env.CAPTURE_GPU_COUNT || '0', 10) || 0;
+  if (envN > 0) return envN;
+  // _gpuCache is declared later (function hoisting makes the ref fine at call
+  // time). null = not probed yet; [] = no GPUs; otherwise one entry per GPU.
+  if (Array.isArray(_gpuCache) && _gpuCache.length > 0) return _gpuCache.length;
+  return 1; // not probed yet / no GPUs → single-device fallback ('all')
 }
 
 // Choose the NVIDIA_VISIBLE_DEVICES value for a capture sidecar. An explicit