diff --git a/build.rs b/build.rs
index e6e6a69197..7bb2f550f8 100644
--- a/build.rs
+++ b/build.rs
@@ -70,6 +70,22 @@ fn main() {
     for dll in delay_load_dlls {
         println!("cargo::rustc-link-arg-bin=rustup-init=/delayload:{dll}.dll");
     }
+    // Also delay load system DLLs that the proxies never call. Every `cargo`,
+    // `rustc` etc. started through rustup is a separate process on Windows, and
+    // each of these DLLs (with what it pulls in, such as gdi32 for user32) costs
+    // start-up time when it is loaded eagerly.
+    //
+    // Builds with the `test` feature (CI's test runs) skip this: they also link
+    // the binary's empty test and bench harnesses, which don't import these
+    // DLLs, so `/WX` would turn the unused `/delayload` into a link error.
+    if env::var_os("CARGO_FEATURE_TEST").is_none() {
+        let delay_load_for_startup = [
+            "user32", "shell32", "ole32", "oleaut32", "crypt32", "secur32", "ws2_32",
+        ];
+        for dll in delay_load_for_startup {
+            println!("cargo::rustc-link-arg-bin=rustup-init=/delayload:{dll}.dll");
+        }
+    }
     // When using delayload, it's necessary to also link delayimp.lib
     // https://learn.microsoft.com/en-us/cpp/build/reference/dependentloadflag?view=msvc-170
     println!("cargo::rustc-link-arg-bin=rustup-init=delayimp.lib");