This makes me rethink my entire architecture, and if that machine is costing me probably at least £50 per month, I could get quite a decent cloud server for that money.

Doing the transcription, that's the only thing that I wouldn't run without the GPU successfully, and the transcription is running really fast, I can't get it to run anywhere near fast enough on CPU, and I'm transcribing about 30 to 40 hours of audio per day, so if I used OpenAI or AssemblyAI to do that,

what would the cost be?

The option is to have a cloud GPU server that I run for 1 hour every 6 hours to catch up on jobs, and it would potentially be even faster than mine,

I mean it would certainly be a lot faster than mine, wouldn't that work well?

So based on my self-hosted solution, and bearing in mind, it's not perfect results, I might try something a bit heavier, but anyway, assuming we use the same models and everything, I currently transcribe every 5 minute audio clip in around 8 to 10 seconds, followed by a 5 minute audio clip in around 8 to 10