Can a given Lambda container process ever be used for multiple concurrent function invocations? If not, then we're dealing with at least 128 MB of RAM per concurrent request. Multithreaded and event-driven servers can do much better than that.