diff --git a/jobs/cloud_controller_ng/spec b/jobs/cloud_controller_ng/spec
index 3bbbe54954..d05ee4f2b2 100644
--- a/jobs/cloud_controller_ng/spec
+++ b/jobs/cloud_controller_ng/spec
@@ -1022,6 +1022,10 @@ properties:
     description: "Configure keep alive timeout for connections to cloud controller. This is a temporary field used for testing."
     default: 100
 
+  cc.nginx_upstream_keepalive_timeout:
+    description: "Keepalive timeout in seconds for nginx upstream connections to Puma. Must be less than cc.puma.persistent_timeout (default 20s) to prevent stale keepalive connections causing broken pipe errors."
+    default: 15
+
   cc.jobs.local.number_of_workers:
     default: 2
     description: "Number of local cloud_controller_worker workers"
diff --git a/jobs/cloud_controller_ng/templates/nginx.conf.erb b/jobs/cloud_controller_ng/templates/nginx.conf.erb
index cf3a2a0c08..2a763d8160 100644
--- a/jobs/cloud_controller_ng/templates/nginx.conf.erb
+++ b/jobs/cloud_controller_ng/templates/nginx.conf.erb
@@ -50,6 +50,8 @@ limit_req_zone $http_authorization zone=<%= zone['name'] %>:10m rate=<%= zone['l
 
   upstream cloud_controller {
     server unix:/var/vcap/data/cloud_controller_ng/cloud_controller.sock;
+    keepalive <%= p("cc.puma.workers") * p("cc.puma.max_threads") %>;
+    keepalive_timeout <%= p("cc.nginx_upstream_keepalive_timeout") %>s;
   }
 
   upstream cloud_controller_metrics {