Fundamentally there's a set of facts you're going to need in memory while you're working on a request, and those data structures should have a similar footprint in any reasonable language. Java is only worse for people who have ignored NIO and are still making blocking calls from huge thread pools (I admit there are a lot of bad client libraries that are not async-first), or people whose workload is so tiny that they don't want to spend code space to gain speed (which the JIT tries hard to do).

Edit: and a deployment this strapped for cash should probably be hot-reloading the services within a single JVM rather than paying for lots of network marshalling across machine boundaries. Their platform has single instances 30x as powerful as they've chosen.