The numbers you stated sound off ($500k capex + electricity per 3 concurrent requests?). Especially now that the frontier has moved to ultra sparse MoE architectures. I’ve also read a couple of commodity inference providers claiming that their unit economics are profitable.