One question is: can they maintain those margins? Anthropic currently is the most expensive provider by far, and it is not clear that Opus delivers enough value to justify the markup over their competitors. E.g. OpenAI charges half of Anthropic's API rates for Sol vs. Opus or Terra vs. Sonnet and appears to be more token-efficient, not to mention all the other providers.

While inference may become more efficient and drive costs down for Anthropic in the long term (though over the next decade, they have hundreds of billions in data center commitments, so it's not going to to happen soon), that goes for every provider. Commoditiziation and competition on price is likely to happen, and with models getting ever better, the value proposition of frontier models would become much less compelling.