diff --git a/Cargo.lock b/Cargo.lock
index b78927f45f..60e7dacc2c 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -715,7 +715,6 @@ dependencies = [
  "criterion 0.5.1",
  "crossterm 0.28.1",
  "futures",
- "num_cpus",
  "parquet",
  "pepita",
  "pollster",
diff --git a/crates/aprender-distribute/Cargo.toml b/crates/aprender-distribute/Cargo.toml
index d7fee9105b..4d41b871d6 100644
--- a/crates/aprender-distribute/Cargo.toml
+++ b/crates/aprender-distribute/Cargo.toml
@@ -14,7 +14,7 @@ documentation = "https://docs.rs/repartir"
 
 [features]
 default = ["cpu"]
-cpu = ["tokio", "num_cpus"]
+cpu = ["tokio"]
 gpu = ["wgpu", "pollster"]
 remote = ["tokio", "bincode", "tracing-subscriber"]
 remote-tls = ["remote", "rustls", "tokio-rustls", "rustls-pemfile"]
@@ -34,7 +34,6 @@ name = "repartir"
 tokio = { version = "1.35", features = ["rt-multi-thread", "sync", "time", "process", "io-util", "macros", "net"], optional = true }
 
 # CPU backend
-num_cpus = { version = "1.16", optional = true }
 
 # GPU backend (pure Rust via gfx-rs)
 # wgpu provides cross-platform GPU compute (Vulkan/Metal/DX12/WebGPU)
diff --git a/crates/aprender-distribute/src/executor/cpu.rs b/crates/aprender-distribute/src/executor/cpu.rs
index 4ba21a3864..548880a6fc 100644
--- a/crates/aprender-distribute/src/executor/cpu.rs
+++ b/crates/aprender-distribute/src/executor/cpu.rs
@@ -36,9 +36,16 @@ impl CpuExecutor {
     /// Creates a new CPU executor.
     ///
     /// Detects the number of available CPU cores automatically.
+    ///
+    /// Uses `std::thread::available_parallelism`, stable since Rust 1.59 and far
+    /// below this workspace's 1.91 MSRV, rather than the `num_cpus` crate. That
+    /// dependency was optional and pulled by ONE feature (`cpu`) while this file
+    /// compiles under any feature enabling `tokio`, so six sibling selections
+    /// could not link it. Falling back to 1 when the count is undeterminable
+    /// matches what `num_cpus::get()` returned in the same situation.
     #[must_use]
     pub fn new() -> Self {
-        let num_cpus = num_cpus::get();
+        let num_cpus = std::thread::available_parallelism().map_or(1, std::num::NonZeroUsize::get);
         info!("CpuExecutor initialized with {num_cpus} cores");
         Self { num_cpus }
     }