A method and system for dynamically altering the delivery of web content to end users based on server load. Responsive to receiving a client request for web content, utilization metrics are collected from a plurality of devices that deliver the web content to the end users. Individual load levels for the devices are determined respectively, based on the utilization metrics of the devices, a combined load level is determined for two or more of the devices having the same device type based on the individual-load levels, and a service level to provide to the client is determined based on the combined load level. The request header is modified to specify a rate to deliver the web content to the client based on the service level. The request is sent with the modified header to one of the devices to serve a variant of the web content to the client at the specified rate.