diff --git a/Cargo.lock b/Cargo.lock
index 17f738ea5..0a3e9f5e0 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -3593,6 +3593,7 @@ dependencies = [
  "lightcraft-merge",
  "lightcraft-raster",
  "lightcraft-sysmem",
+ "lightcraft-testkit",
  "lightcraft-ui-egui",
  "serde_json",
 ]
diff --git a/apps/lightcraft-cli/Cargo.toml b/apps/lightcraft-cli/Cargo.toml
index b82aca1d5..e55d458b3 100644
--- a/apps/lightcraft-cli/Cargo.toml
+++ b/apps/lightcraft-cli/Cargo.toml
@@ -25,6 +25,9 @@ lightcraft-merge = { workspace = true }
 serde_json = { workspace = true }
 dhat = { workspace = true, optional = true }
 
+[dev-dependencies]
+lightcraft-testkit = { workspace = true }
+
 [features]
 default = ["denoise"]
 denoise = ["lightcraft-engine/denoise"]
diff --git a/apps/lightcraft-cli/tests/it/cli.rs b/apps/lightcraft-cli/tests/it/cli.rs
index 8a11f5c9a..70f7143e6 100644
--- a/apps/lightcraft-cli/tests/it/cli.rs
+++ b/apps/lightcraft-cli/tests/it/cli.rs
@@ -443,7 +443,12 @@ fn snapshot_defaults_to_cpu_without_gpu_environment_overrides() {
     let script = tmp("default-cpu.jsonl");
     std::fs::write(
         &script,
-        format!("{}\n{}\n", json!({"method": "engine.execute", "params": {"command": "app.gpu"}}), json!({"method": "ui.screenshot"})),
+        format!(
+            "{}\n{}\n{}\n",
+            json!({"method": "engine.execute", "params": {"command": "app.gpu", "params": {"wait": false}}}),
+            json!({"method": "engine.execute", "params": {"command": "app.gpu"}}),
+            json!({"method": "ui.screenshot"})
+        ),
     )
     .unwrap();
     let image = tmp("default-cpu.png");
@@ -457,15 +462,52 @@ fn snapshot_defaults_to_cpu_without_gpu_environment_overrides() {
     assert!(output.status.success(), "status {}: {}", output.status, String::from_utf8_lossy(&output.stderr));
     let replies: Vec<Value> = String::from_utf8_lossy(&output.stdout).lines().map(|line| serde_json::from_str(line).unwrap()).collect();
     assert!(replies.iter().all(|reply| reply["ok"] == true), "{replies:?}");
-    let gpu = &replies[0]["result"];
-    assert_eq!(gpu["enabled"], false, "{gpu}");
-    assert_eq!(gpu["available"], false, "{gpu}");
-    assert_eq!(gpu["adapter"], Value::Null, "no device is created: {gpu}");
-    assert!(gpu["reason"].as_str().is_some_and(|reason| reason.contains("preference")), "{gpu}");
+    // `wait: false` (what the UI sends) still says why the GPU is off: no device is being created
+    for gpu in [&replies[0]["result"], &replies[1]["result"]] {
+        assert_eq!(gpu["enabled"], false, "{gpu}");
+        assert_eq!(gpu["available"], false, "{gpu}");
+        assert_eq!(gpu["adapter"], Value::Null, "no device is created: {gpu}");
+        assert!(gpu["reason"].as_str().is_some_and(|reason| reason.contains("preference")), "{gpu}");
+    }
     let decoded = lightcraft_codecs::decode(&std::fs::read(image).unwrap(), Default::default()).unwrap();
     assert_eq!((decoded.width, decoded.height), (480, 320));
 }
 
+/// Issue #136: `app.gpu {enabled: true, wait: false}` (what the UI sends on its first frame) answers
+/// at once while the device is still being created: enabled, and nothing known yet (`null`), instead
+/// of waiting for the driver. It starts creating the device (off the calling thread), so this test
+/// touches the GPU driver and skips when the environment turns the GPU off.
+#[test]
+fn enabling_the_gpu_without_waiting_answers_before_the_device_exists() {
+    if std::env::var("LIGHTCRAFT_GPU").is_ok_and(|v| matches!(v.trim(), "0" | "off" | "false" | "no")) {
+        lightcraft_testkit::skip!(Gpu, "LIGHTCRAFT_GPU=0: enabling the GPU starts creating a device");
+        return;
+    }
+    let script = tmp("gpu-starting.jsonl");
+    std::fs::write(
+        &script,
+        format!("{}\n", json!({"method": "engine.execute", "params": {"command": "app.gpu", "params": {"enabled": true, "wait": false}}})),
+    )
+    .unwrap();
+    let image = tmp("gpu-starting.png");
+    let output = Command::new(binary())
+        .env_remove("LIGHTCRAFT_GPU")
+        .env_remove("LIGHTCRAFT_GPU_BACKEND")
+        .env_remove("WGPU_BACKEND")
+        .args(["snapshot", "--demo", "--script", script.to_str().unwrap(), "-o", image.to_str().unwrap(), "--size", "480x320"])
+        .output()
+        .unwrap();
+    assert!(output.status.success(), "status {}: {}", output.status, String::from_utf8_lossy(&output.stderr));
+    let replies: Vec<Value> = String::from_utf8_lossy(&output.stdout).lines().map(|line| serde_json::from_str(line).unwrap()).collect();
+    assert_eq!(replies.len(), 1, "{replies:?}");
+    assert_eq!(replies[0]["ok"], true, "{replies:?}");
+    let gpu = &replies[0]["result"];
+    assert_eq!(gpu["enabled"], true, "{gpu}");
+    for field in ["available", "adapter", "reason"] {
+        assert_eq!(gpu[field], Value::Null, "{field} is not known while the device is created: {gpu}");
+    }
+}
+
 /// Issue #136: with the GPU switched off from the environment the UI starts and renders on the CPU,
 /// without creating a GPU device (no driver is loaded), and says why.
 #[test]
diff --git a/crates/engine/src/cmd/query.rs b/crates/engine/src/cmd/query.rs
index a3926bc51..469253805 100644
--- a/crates/engine/src/cmd/query.rs
+++ b/crates/engine/src/cmd/query.rs
@@ -230,10 +230,29 @@ pub fn specs() -> Vec<CommandSpec> {
                 "versions": ph.versions.iter().map(|v| json!({"name": v.name, "created": v.created})).collect::<Vec<_>>(),
             }))
         }),
-        cmd!(query "app.gpu", "GPU Rendering", [], None, "{enabled?: bool} — allow/forbid GPU rendering (CPU fallback; LIGHTCRAFT_GPU=0 forbids it for the process); returns {enabled, available, adapter, reason (why the GPU is off), lastFallback (latest render redone on the CPU, and why)}", always, |_, p| {
+        cmd!(query "app.gpu", "GPU Rendering", [], None, "{enabled?: bool, wait?: bool} — allow/forbid GPU rendering (CPU fallback; LIGHTCRAFT_GPU=0 forbids it for the process); returns {enabled, available, adapter, reason (why the GPU is off), lastFallback (latest render redone on the CPU, and why)}. wait: false never waits for a device still being created (available, adapter and reason are then null and the device is created in the background; with GPU rendering off they say so at once)", always, |_, p| {
             if let Some(on) = p.get("enabled").and_then(Value::as_bool) {
                 lightcraft_gpu::set_enabled(on);
             }
+            // the UI applies the setting on its first frame: it must not wait seconds for the
+            // device there (issue #136; `warm_up` creates it off the UI thread)
+            if p.get("wait").and_then(Value::as_bool) == Some(false) && !lightcraft_gpu::ready() {
+                // off (preference, environment): known at once, and asking the adapter would
+                // create the device; on: still being created, so nothing is known yet
+                let on = lightcraft_gpu::enabled();
+                if on {
+                    // a caller that never waits still gets an answer later (off the UI thread;
+                    // starts at most once)
+                    lightcraft_gpu::warm_up();
+                }
+                return Ok(json!({
+                    "enabled": on,
+                    "available": if on { Value::Null } else { Value::Bool(false) },
+                    "adapter": null,
+                    "reason": if on { None } else { lightcraft_gpu::unavailable_reason() },
+                    "lastFallback": lightcraft_gpu::last_fallback(),
+                }));
+            }
             Ok(json!({
                 "enabled": lightcraft_gpu::enabled(),
                 "available": lightcraft_gpu::available(),
diff --git a/crates/gpu/src/lib.rs b/crates/gpu/src/lib.rs
index 452e2dd67..e66a8fadd 100644
--- a/crates/gpu/src/lib.rs
+++ b/crates/gpu/src/lib.rs
@@ -193,7 +193,9 @@ pub fn reset_failures() {
 /// Allow or forbid GPU rendering at runtime (a preference). `LIGHTCRAFT_GPU=0` forbids it for the
 /// whole process regardless.
 pub fn set_enabled(on: bool) {
-    ENABLED.store(on, Ordering::Relaxed);
+    // (SeqCst, as `warm_up`'s thread reads it after releasing its start: a switch back on is
+    // never missed there)
+    ENABLED.store(on, Ordering::SeqCst);
 }
 
 /// Whether GPU rendering is allowed (environment, preference, no earlier failure — real, or
@@ -308,13 +310,33 @@ pub(crate) fn existing_device() -> Option<&'static ctx::Gpu> {
 
 /// Create the device and compile the kernels on a background thread now (once the window is up),
 /// so the first render doesn't wait ~0.3–0.4 s for it and the UI thread never does. Does nothing
-/// while GPU rendering is turned off.
+/// while GPU rendering is turned off. Repeated calls start one thread at most.
 pub fn warm_up() {
     #[cfg(not(target_arch = "wasm32"))]
-    if !env_disabled() && ENABLED.load(Ordering::Relaxed) && GPU.get().is_none() {
-        let _ = std::thread::Builder::new().name("lc-gpu-init".into()).spawn(|| {
-            let _ = device();
+    if !env_disabled() && ENABLED.load(Ordering::SeqCst) && GPU.get().is_none() {
+        static STARTED: AtomicBool = AtomicBool::new(false);
+        if STARTED.swap(true, Ordering::SeqCst) {
+            return;
+        }
+        let spawned = std::thread::Builder::new().name("lc-gpu-init".into()).spawn(|| {
+            loop {
+                let _ = device();
+                if GPU.get().is_some() || shutting_down() || env_disabled() {
+                    break;
+                }
+                // turned off before the thread ran (nothing created): a later call may start
+                // again. Released first, then the preference read again: a call made while this
+                // thread still held the flag (turned back on meanwhile) returned without starting
+                // one, so this thread tries again itself rather than leave "starting" for good.
+                STARTED.store(false, Ordering::SeqCst);
+                if !ENABLED.load(Ordering::SeqCst) || STARTED.swap(true, Ordering::SeqCst) {
+                    break;
+                }
+            }
         });
+        if spawned.is_err() {
+            STARTED.store(false, Ordering::SeqCst);
+        }
     }
 }
 
diff --git a/crates/ui-egui/src/app/prefs.rs b/crates/ui-egui/src/app/prefs.rs
index 3dcafeb14..9a5c612f1 100644
--- a/crates/ui-egui/src/app/prefs.rs
+++ b/crates/ui-egui/src/app/prefs.rs
@@ -34,7 +34,7 @@ impl LightcraftApp {
     pub(super) fn apply_settings(&mut self, ctx: &egui::Context) {
         if self.gpu_applied != Some(self.ui.settings.gpu) {
             self.gpu_applied = Some(self.ui.settings.gpu);
-            let _ = self.session.execute("app.gpu", &serde_json::json!({"enabled": self.ui.settings.gpu}));
+            let _ = self.session.execute("app.gpu", &serde_json::json!({"enabled": self.ui.settings.gpu, "wait": false}));
             // GPU device + kernels off the UI thread, once the window is up and only when GPU
             // rendering is on: a broken driver must not keep the window from appearing (issue #136)
             if self.ui.settings.gpu {
diff --git a/crates/ui-egui/src/dialogs/settings/performance.rs b/crates/ui-egui/src/dialogs/settings/performance.rs
index 88ff55f6e..370cf08ba 100644
--- a/crates/ui-egui/src/dialogs/settings/performance.rs
+++ b/crates/ui-egui/src/dialogs/settings/performance.rs
@@ -11,7 +11,15 @@ pub(super) fn performance_tab(app: &mut LightcraftApp, ui: &mut egui::Ui, t: &To
     #[cfg(target_os = "windows")]
     check(ui, "settings.composeWindow", &mut app.ui.settings.compose_window, "Prevent G-SYNC/FreeSync flicker (applies at next launch)");
     check(ui, "settings.gpu", &mut app.ui.settings.gpu, "Use the GPU for rendering");
-    let status = if !gpu::available() {
+    // `available` would wait for a device still being created (issue #136)
+    let status = if gpu::enabled() && !gpu::ready() {
+        // the device is created on another thread, which doesn't wake the UI when it's done;
+        // make sure that thread runs (cheap and idempotent: a start lost to an off/on switch
+        // can't leave "starting" for good)
+        gpu::warm_up();
+        ui.ctx().request_repaint_after(std::time::Duration::from_millis(250));
+        "GPU: starting…".to_string()
+    } else if !gpu::available() {
         match gpu::unavailable_reason() {
             Some(why) => crate::i18n::tr_format!("Rendering on the CPU: {why}", why = why),
             None => "No usable GPU found: rendering on the CPU".to_string(),