I am pretty sure OpenAI and Claude actually have a higher hardware cost per request that Microsoft. Github is actually most just a standard REST API app, although with some caching and git specific operations in the background. It is not as hardware intensive as GPU-based inference per quest with huge contexts and low latency.

Also Github's feature offering has been incredibly stable for the last 6 years. There isn't an excuse for it to be unstable like this.