Reading about this from what sources I can find it's actually more ambiguous whether or not they are profitable . It looks like they are profitable per token if you ignore the continuing R&D and capex for the next model . So they will probably be able to operate unsubsidized if they get their costs low enough some will survive. Chatgpt sol just dropped with relatively high efficiency per unit of intelligence while they were getting smoked by anthropic in real world utility but by anthropic really getting intelligence through much more brute force and thus much more extreme cost per token , which shows people will pay high costs per token for the utility they get from it , but at some point you have to compete on efficiency and input costs . The Chinese models are really efficient. If they keep dropping open source weights by 2030 we will surely have open source models as good as mythos which I would say is good enough for 99% of people. After that its pure commodity inference purchase x efficiency of the model . Maybe not a capitalist goldmine, but the end result is useful and if we get enough capacity expansion in chips and memory the deficit will turn into a glut and mythos grade home systems will be viable for power users a few years after that