If they are profitable on inference currently, they will be able to recover. People in SWE will continue paying for current SOTA models even if better models are not achieved.
Alternatively, latecomers will make use of newly cheaply available compute (from the firesales of failing companies) to produce models that match their quality, while having to invest only a fraction of what the first wave had to, allowing them to push the price below the cost floor of the first wave and making them go under.