Comment by skohan

I think hardware scarcity plays a part in inference margins. They would get squeezed super quickly if anyone who wanted could just build an inference cluster, but that's not currently the case.