diff --git a/axum/CHANGELOG.md b/axum/CHANGELOG.md
index 77064e03..8ffdbe6f 100644
--- a/axum/CHANGELOG.md
+++ b/axum/CHANGELOG.md
@@ -73,6 +73,9 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
   `IntoResponseParts` so `([("x-foo", "foo")], response)` now works ([#797])
 - **breaking:** `InvalidJsonBody` has been replaced with `JsonDataError` to clearly signal that the
   request body was syntactically valid JSON but couldn't be deserialized into the target type
+- **breaking:** `Handler` is no longer an `#[async_trait]` but instead has an
+  associated `Future` type. That allows users to build their own `Handler` types
+  without paying the cost of `#[async_trait]` ([#879])
 - **changed:** New `JsonSyntaxError` variant added to `JsonRejection`. This is returned when the
   request body contains syntactically invalid JSON
 - **fixed:** Set `Allow` header when responding with `405 Method Not Allowed` ([#733])
@@ -104,6 +107,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 [#827]: https://github.com/tokio-rs/axum/pull/827
 [#841]: https://github.com/tokio-rs/axum/pull/841
 [#842]: https://github.com/tokio-rs/axum/pull/842
+[#879]: https://github.com/tokio-rs/axum/pull/879
 
 # 0.4.4 (13. January, 2022)
 
diff --git a/axum/src/handler/future.rs b/axum/src/handler/future.rs
index 6bd39b35..b8b3e4e8 100644
--- a/axum/src/handler/future.rs
+++ b/axum/src/handler/future.rs
@@ -1,14 +1,52 @@
 //! Handler future types.
 
 use crate::response::Response;
-use futures_util::future::{BoxFuture, Map};
-use std::convert::Infallible;
+use futures_util::future::Map;
+use http::Request;
+use pin_project_lite::pin_project;
+use std::{convert::Infallible, future::Future, pin::Pin, task::Context};
+use tower::util::Oneshot;
+use tower_service::Service;
 
 opaque_future! {
     /// The response future for [`IntoService`](super::IntoService).
-    pub type IntoServiceFuture =
+    pub type IntoServiceFuture<F> =
         Map<
-            BoxFuture<'static, Response>,
+            F,
             fn(Response) -> Result<Response, Infallible>,
         >;
 }
+
+pin_project! {
+    /// The response future for [`Layered`](super::Layered).
+    pub struct LayeredFuture<S, ReqBody>
+    where
+        S: Service<Request<ReqBody>>,
+    {
+        #[pin]
+        inner: Map<Oneshot<S, Request<ReqBody>>, fn(Result<S::Response, S::Error>) -> Response>,
+    }
+}
+
+impl<S, ReqBody> LayeredFuture<S, ReqBody>
+where
+    S: Service<Request<ReqBody>>,
+{
+    pub(super) fn new(
+        inner: Map<Oneshot<S, Request<ReqBody>>, fn(Result<S::Response, S::Error>) -> Response>,
+    ) -> Self {
+        Self { inner }
+    }
+}
+
+impl<S, ReqBody> Future for LayeredFuture<S, ReqBody>
+where
+    S: Service<Request<ReqBody>>,
+{
+    type Output = Response;
+
+    #[inline]
+    fn poll(self: Pin<&mut Self>, cx: &mut Context<'_>) -> std::task::Poll<Self::Output> {
+        self.project().inner.poll(cx)
+    }
+}
diff --git a/axum/src/handler/into_service.rs b/axum/src/handler/into_service.rs
index 1cacfa20..34f36b2d 100644
--- a/axum/src/handler/into_service.rs
+++ b/axum/src/handler/into_service.rs
@@ -60,7 +60,7 @@ where
 {
     type Response = Response;
     type Error = Infallible;
-    type Future = super::future::IntoServiceFuture;
+    type Future = super::future::IntoServiceFuture<H::Future>;
 
     #[inline]
     fn poll_ready(&mut self, _cx: &mut Context<'_>) -> Poll<Result<(), Self::Error>> {
@@ -74,7 +74,8 @@ where
         use futures_util::future::FutureExt;
 
         let handler = self.handler.clone();
-        let future = Handler::call(handler, req).map(Ok::<_, Infallible> as _);
+        let future = Handler::call(handler, req);
+        let future = future.map(Ok as _);
 
         super::future::IntoServiceFuture::new(future)
     }
diff --git a/axum/src/handler/mod.rs b/axum/src/handler/mod.rs
index fc58831c..5818a70a 100644
--- a/axum/src/handler/mod.rs
+++ b/axum/src/handler/mod.rs
@@ -80,9 +80,8 @@ use crate::{
     routing::IntoMakeService,
     BoxError,
 };
-use async_trait::async_trait;
 use http::Request;
-use std::{fmt, future::Future, marker::PhantomData};
+use std::{fmt, future::Future, marker::PhantomData, pin::Pin};
 use tower::ServiceExt;
 use tower_layer::Layer;
 use tower_service::Service;
@@ -98,10 +97,12 @@ pub use self::into_service::IntoService;
 /// implemented to closures of the right types.
 ///
 /// See the [module docs](crate::handler) for more details.
-#[async_trait]
 pub trait Handler<T, B = Body>: Clone + Send + Sized + 'static {
+    /// The type of future calling this handler returns.
+    type Future: Future<Output = Response> + Send + 'static;
+
     /// Call the handler with the given request.
-    async fn call(self, req: Request<B>) -> Response;
+    fn call(self, req: Request<B>) -> Self::Future;
 
     /// Apply a [`tower::Layer`] to the handler.
     ///
@@ -247,7 +248,6 @@ pub trait Handler<T, B = Body>: Clone + Send + Sized + 'static {
     }
 }
 
-#[async_trait]
 impl<F, Fut, Res, B> Handler<(), B> for F
 where
     F: FnOnce() -> Fut + Clone + Send + 'static,
@@ -255,14 +255,15 @@ where
     Res: IntoResponse,
     B: Send + 'static,
 {
-    async fn call(self, _req: Request<B>) -> Response {
-        self().await.into_response()
+    type Future = Pin<Box<dyn Future<Output = Response> + Send>>;
+
+    fn call(self, _req: Request<B>) -> Self::Future {
+        Box::pin(async move { self().await.into_response() })
     }
 }
 
 macro_rules! impl_handler {
     ( $($ty:ident),* $(,)? ) => {
-        #[async_trait]
         #[allow(non_snake_case)]
         impl<F, Fut, B, Res, $($ty,)*> Handler<($($ty,)*), B> for F
         where
@@ -272,19 +273,23 @@ macro_rules! impl_handler {
             Res: IntoResponse,
             $( $ty: FromRequest<B> + Send,)*
         {
-            async fn call(self, req: Request<B>) -> Response {
-                let mut req = RequestParts::new(req);
+            type Future = Pin<Box<dyn Future<Output = Response> + Send>>;
 
-                $(
-                    let $ty = match $ty::from_request(&mut req).await {
-                        Ok(value) => value,
-                        Err(rejection) => return rejection.into_response(),
-                    };
-                )*
+            fn call(self, req: Request<B>) -> Self::Future {
+                Box::pin(async move {
+                    let mut req = RequestParts::new(req);
 
-                let res = self($($ty,)*).await;
+                    $(
+                        let $ty = match $ty::from_request(&mut req).await {
+                            Ok(value) => value,
+                            Err(rejection) => return rejection.into_response(),
+                        };
+                    )*
 
-                res.into_response()
+                    let res = self($($ty,)*).await;
+
+                    res.into_response()
+                })
             }
         }
     };
@@ -318,7 +323,6 @@ where
     }
 }
 
-#[async_trait]
 impl<S, T, ReqBody, ResBody> Handler<T, ReqBody> for Layered<S, T>
 where
     S: Service<Request<ReqBody>, Response = Response<ResBody>> + Clone + Send + 'static,
@@ -329,11 +333,18 @@ where
     ResBody: HttpBody<Data = Bytes> + Send + 'static,
     ResBody::Error: Into<BoxError>,
 {
-    async fn call(self, req: Request<ReqBody>) -> Response {
-        match self.svc.oneshot(req).await {
-            Ok(res) => res.map(boxed),
-            Err(res) => res.into_response(),
-        }
+    type Future = future::LayeredFuture<S, ReqBody>;
+
+    fn call(self, req: Request<ReqBody>) -> Self::Future {
+        use futures_util::future::{FutureExt, Map};
+
+        let future: Map<_, fn(Result<S::Response, S::Error>) -> _> =
+            self.svc.oneshot(req).map(|result| match result {
+                Ok(res) => res.map(boxed),
+                Err(res) => res.into_response(),
+            });
+
+        future::LayeredFuture::new(future)
     }
 }