diff --git a/socktop_agent/src/metrics.rs b/socktop_agent/src/metrics.rs
index 52b1858..e1e8c4e 100644
--- a/socktop_agent/src/metrics.rs
+++ b/socktop_agent/src/metrics.rs
@@ -153,20 +153,41 @@ pub async fn collect_fast_metrics(state: &AppState) -> Metrics {
             .collect()
     };
 
-    // GPUs: refresh only when cache is stale
-    let gpus = if cached_gpus().is_some() {
-        cached_gpus()
-    } else if gpu_enabled() {
-        let v = match collect_all_gpus() {
-            Ok(v) if !v.is_empty() => Some(v),
-            Ok(_) => None,
-            Err(e) => {
-                warn!("gpu collection failed: {e}");
-                None
+    // GPUs: if we already determined none exist, short-circuit (no repeated probing)
+    let gpus = if gpu_enabled() {
+        if state.gpu_checked.load(std::sync::atomic::Ordering::Acquire)
+            && !state.gpu_present.load(std::sync::atomic::Ordering::Relaxed)
+        {
+            None
+        } else if cached_gpus().is_some() {
+            cached_gpus()
+        } else {
+            let v = match collect_all_gpus() {
+                Ok(v) if !v.is_empty() => Some(v),
+                Ok(_) => None,
+                Err(e) => {
+                    warn!("gpu collection failed: {e}");
+                    None
+                }
+            };
+            // First probe records presence; subsequent calls rely on cache flags.
+            if !state
+                .gpu_checked
+                .swap(true, std::sync::atomic::Ordering::AcqRel)
+            {
+                if v.is_some() {
+                    state
+                        .gpu_present
+                        .store(true, std::sync::atomic::Ordering::Release);
+                } else {
+                    state
+                        .gpu_present
+                        .store(false, std::sync::atomic::Ordering::Release);
+                }
             }
-        };
-        set_gpus(v.clone());
-        v
+            set_gpus(v.clone());
+            v
+        }
     } else {
         None
     };
diff --git a/socktop_agent/src/state.rs b/socktop_agent/src/state.rs
index 53edd32..9b8a6ce 100644
--- a/socktop_agent/src/state.rs
+++ b/socktop_agent/src/state.rs
@@ -2,7 +2,7 @@
 
 #[cfg(target_os = "linux")]
 use std::collections::HashMap;
-use std::sync::atomic::AtomicUsize;
+use std::sync::atomic::{AtomicBool, AtomicUsize};
 use std::sync::Arc;
 use sysinfo::{Components, Disks, Networks, System};
 use tokio::sync::Mutex;
@@ -34,6 +34,9 @@ pub struct AppState {
     pub client_count: Arc<AtomicUsize>,
 
     pub auth_token: Option<String>,
+    // GPU negative cache (probe once). gpu_checked=true after first attempt; gpu_present reflects result.
+    pub gpu_checked: Arc<AtomicBool>,
+    pub gpu_present: Arc<AtomicBool>,
 }
 
 impl AppState {
@@ -54,6 +57,8 @@ impl AppState {
             auth_token: std::env::var("SOCKTOP_TOKEN")
                 .ok()
                 .filter(|s| !s.is_empty()),
+            gpu_checked: Arc::new(AtomicBool::new(false)),
+            gpu_present: Arc::new(AtomicBool::new(false)),
         }
     }
 }