@timkellogg.me would be interesting to see how the economies of scale work out on how much cheaper inference (per token costs) is on datacenter machines vs this