diff --git a/wicket-common/src/lib.rs b/wicket-common/src/lib.rs
index 9e92d20c0a..aea0634ce7 100644
--- a/wicket-common/src/lib.rs
+++ b/wicket-common/src/lib.rs
@@ -4,6 +4,13 @@
 
 // Copyright 2023 Oxide Computer Company
 
+use std::time::Duration;
+
 pub mod rack_setup;
 pub mod rack_update;
 pub mod update_events;
+
+// WICKETD_TIMEOUT used to be 1 second, but that might be too short (and in
+// particular might be responsible for
+// https://github.com/oxidecomputer/omicron/issues/3103).
+pub const WICKETD_TIMEOUT: Duration = Duration::from_secs(5);
diff --git a/wicket/src/cli/rack_update.rs b/wicket/src/cli/rack_update.rs
index f539c22c35..fa41fa7b8c 100644
--- a/wicket/src/cli/rack_update.rs
+++ b/wicket/src/cli/rack_update.rs
@@ -22,6 +22,7 @@ use update_engine::{
 };
 use wicket_common::{
     rack_update::ClearUpdateStateResponse, update_events::EventReport,
+    WICKETD_TIMEOUT,
 };
 use wicketd_client::types::{ClearUpdateStateParams, StartUpdateParams};
 
@@ -31,7 +32,7 @@ use crate::{
         parse_event_report_map, ComponentId, CreateClearUpdateStateOptions,
         CreateStartUpdateOptions,
     },
-    wicketd::{create_wicketd_client, WICKETD_TIMEOUT},
+    wicketd::create_wicketd_client,
 };
 
 use super::command::CommandOutput;
diff --git a/wicket/src/wicketd.rs b/wicket/src/wicketd.rs
index ec1130a594..a951bf428b 100644
--- a/wicket/src/wicketd.rs
+++ b/wicket/src/wicketd.rs
@@ -10,6 +10,7 @@ use std::net::SocketAddrV6;
 use tokio::sync::mpsc::{self, Sender, UnboundedSender};
 use tokio::time::{interval, Duration, MissedTickBehavior};
 use wicket_common::rack_update::{SpIdentifier, SpType};
+use wicket_common::WICKETD_TIMEOUT;
 use wicketd_client::types::{
     AbortUpdateOptions, ClearUpdateStateOptions, ClearUpdateStateParams,
     GetInventoryParams, GetInventoryResponse, GetLocationResponse,
@@ -38,10 +39,6 @@ impl From<ComponentId> for SpIdentifier {
 }
 
 const WICKETD_POLL_INTERVAL: Duration = Duration::from_millis(500);
-// WICKETD_TIMEOUT used to be 1 second, but that might be too short (and in
-// particular might be responsible for
-// https://github.com/oxidecomputer/omicron/issues/3103).
-pub(crate) const WICKETD_TIMEOUT: Duration = Duration::from_secs(5);
 
 // Assume that these requests are periodic on the order of seconds or the
 // result of human interaction. In either case, this buffer should be plenty
diff --git a/wicketd/src/http_entrypoints.rs b/wicketd/src/http_entrypoints.rs
index d6cb6ebd6d..cff0436dac 100644
--- a/wicketd/src/http_entrypoints.rs
+++ b/wicketd/src/http_entrypoints.rs
@@ -51,6 +51,7 @@ use std::time::Duration;
 use tokio::io::AsyncWriteExt;
 use wicket_common::rack_setup::PutRssUserConfigInsensitive;
 use wicket_common::update_events::EventReport;
+use wicket_common::WICKETD_TIMEOUT;
 
 use crate::ServerContext;
 
@@ -896,21 +897,37 @@ async fn post_start_update(
     // 1. We haven't pulled its state in our inventory (most likely cause: the
     //    cubby is empty; less likely cause: the SP is misbehaving, which will
     //    make updating it very unlikely to work anyway)
-    // 2. We have pulled its state but our hardware manager says we can't update
-    //    it (most likely cause: the target is the sled we're currently running
-    //    on; less likely cause: our hardware manager failed to get our local
-    //    identifying information, and it refuses to update this target out of
-    //    an abundance of caution).
+    // 2. We have pulled its state but our hardware manager says we can't
+    //    update it (most likely cause: the target is the sled we're currently
+    //    running on; less likely cause: our hardware manager failed to get our
+    //    local identifying information, and it refuses to update this target
+    //    out of an abundance of caution).
     //
-    // First, get our most-recently-cached inventory view.
-    let inventory = match rqctx.mgs_handle.get_cached_inventory().await {
-        Ok(inventory) => inventory,
-        Err(ShutdownInProgress) => {
+    // First, get our most-recently-cached inventory view. (Only wait 80% of
+    // WICKETD_TIMEOUT for this -- if the inventory isn't available, then we
+    // should produce a useful error message rather than timing out on the
+    // client.)
+    let inventory = match tokio::time::timeout(
+        WICKETD_TIMEOUT.mul_f32(0.8),
+        rqctx.mgs_handle.get_cached_inventory(),
+    )
+    .await
+    {
+        Ok(Ok(inventory)) => inventory,
+        Ok(Err(ShutdownInProgress)) => {
             return Err(HttpError::for_unavail(
                 None,
                 "Server is shutting down".into(),
             ));
         }
+        Err(_) => {
+            // Use 400 Bad Request instead of 503 Service Unavailable so that
+            // the error message is propagated to the client.
+            return Err(HttpError::for_bad_request(
+                None,
+                "Rack inventory not yet available (is MGS alive?)".into(),
+            ));
+        }
     };
 
     // Error cases.