Many webservers allow you to serve a compressed file (stored on disk) and _decompress_ when a client specifically can't support the compressed encoding. Since most clients should support compression, this means you only use the CPU for the rarer case where the uncompressed data is required.
But what if an analyst needs to access this data and run their regressions on a potato? Surely that use case is worth adding a few libraries to handle.
hmm good point, can you draft up an architecture plan using multiple microservices and redundancy via a kubernetes cluster and have it on my desk by Monday please.
Eg, http://nginx.org/en/docs/http/ngx_http_gunzip_module.html