It does seem like the open-weight models don't have easily monetizable adjacent money streams like support and services, best they have is inference cost. But the whole point of the open weights is that anyone else can come and take that part from them. It's a tough place to be in if the goal is only profitability. But since AI was declared a national security thing here in the US, the strategic value of ensuring we don't end up dominating global AI is probably worth it. And a steal at the 500mil per year loss they claim.
I do wonder what sustainable future looks like for an open weights lab, once we're past this wackiness. A very well tuned harness? That's kind of its own product. On-prem seployment support?
This is just standard "ram/chips are so expensive now that all consumer devices had to raise prices and thus sold less". With maybe a little general reduction in overall consumer spending?