One shared vCPU on Vercel's free tier, which is what this page costs to run and not what the parser costs. The same model reads a name in about 2,400 µs through the Python binding on a desktop Ryzen, and far less again from C++ in a batch, where the work is shared across names and threads. Round trip is network and JSON on top.
One release name per line. Copy a listing straight off a torrent site and paste it in as it comes: the lines that are not release names are filtered out before anything is parsed.