The typical playbook for a VC funded startup is to race to a monopoly where the company can have higher margins. Prices continuing to go down for the consumer over time would require competition to stay high in the long term, and even then it’s not clear if even current prices are profitable.

The current level of AI has plenty of inherent competition from local models. In the long term, most of the profit will probably be from very smart models that run at something closer to datacenter scale over long inference loops - where local inference can't do much and even third-party inference/small neoclouds will be severely challenged. That is a very natural "moat" and has natural cross-efficiencies with AI model training, which requires a similar scale.