> I think that it's not only possible but likely that the frontier labs are profitable on an ongoing basis, and it's mainly the research and capex which make them look like impending disasters.

And that is the much more relevant problem. Capable open-weight models from China, European labs or competitors with cash to burn (i.e. Meta!) erode all the moat that the major labs have, and tooling functionality like MCP, CLI interfaces or integration into third-party tools like IDEs quickly gets cloned.

Simply printing money by running inference does not justify valuations in the billions of dollars, because anyone can do so.