Writing to a stream is the way to go, as it will rougly consume not more than the current "record" and associated memory. That stream can be a FileStream (if you create a file) or the ASP.NET stream (if you write directly to the web), or any other useful stream.

The advantage of creating a file (using the FileStream) is to be able to cache the data to serve the same request over and over. Depending on your need, this can be a real benefit. You must come up with an intelligent algorithm to determine the file path & name from the input. This will be the cache key. Once you have a file, you can use the TransmitFile api which leverages Windows kernel cache and in general very efficient. You can also play with HTTP client caches (headers like last-modified-since, etc.), so next time the client request the same information, you may return a not modified (HTTP 304 status code) response. The disadvantages of using cache files is you will need to manage these files, disk space, expiration, etc.

Now, Linq or IDataReader should not change much about perf or memory consumption provided you don't use Linq method that materialize the whole data (exhaust the stream) or a big part of it. That means you will need to avoid ToArray(), ToList() methods and other methods like this, and concentrate only on "streamed" methods (enumerations, skips, while, etc.).