I might be wrong, but I believe the THREDDS server the article points to (by way of https://cds.nccs.nasa.gov/nex-gddp/) supports OPeNDAP, which means you can pull subsets of the data, i.e. if you've got a 40GB file, you could slice a tiny portion of that off for downloading.
I work with the THREDDS team at Unidata. Though I still need to study this particular dataset, in general yes, the THREDDS server can do serverside subsetting of the data so that reasonably sized data is shipped over the wire.
