An actual Xeon is not going to draw its TDP in power under a simple workload like serving static files. Even if you get it up to 100% CPU utilization, it will probably not be drawing its TDP. Modern intel processors have a bunch of mechanisms built in for managing power draw (and similarly, heat output) - they can clock up and down in response to workloads, bringing a single core up above standard clock to speed up single-threaded loads and bringing all the cores down when the machine is doing less intensive things (like running a message pump or waiting for socket connections).
If anything, I'd expect static file serving on a Xeon to produce no more than say 40% of TDP. If you're lucky, serving up all the static files will load all the cores fairly evenly and get the CPU close to '100%', but none of the floating point or integer logic units will be remotely loaded - it'll be almost exclusively branch/copy work, which isn't going to put much load on the CPU itself or draw much power or generate much heat. It's also going to be spending tons of time waiting (on the NIC, etc) instead of actually doing computation, which can generate a lot less heat if the waits are done using the modern busy wait instructions instead of a spin loop.
EDIT: A comment in the OP provides a conservative estimate of 43W for the actual draw of the entire xeon-based system (not just the CPU) in the benchmark. He also points out that it has more RAM (which will increase power draw).
Serving static files is a kind of thing that could be done with very little CPU involvement. Once the file is cached (or memory-mapped), just point the NIC processor to it and tell it to pipe the memory block through the network connection and head off to nobler jobs.
If anything, I'd expect static file serving on a Xeon to produce no more than say 40% of TDP. If you're lucky, serving up all the static files will load all the cores fairly evenly and get the CPU close to '100%', but none of the floating point or integer logic units will be remotely loaded - it'll be almost exclusively branch/copy work, which isn't going to put much load on the CPU itself or draw much power or generate much heat. It's also going to be spending tons of time waiting (on the NIC, etc) instead of actually doing computation, which can generate a lot less heat if the waits are done using the modern busy wait instructions instead of a spin loop.
EDIT: A comment in the OP provides a conservative estimate of 43W for the actual draw of the entire xeon-based system (not just the CPU) in the benchmark. He also points out that it has more RAM (which will increase power draw).