Will accounting firms eventually need to run their own local/specialized AI models instead of relying on big frontier models like GPT or Claude?

The view expressed is that generative frontier models (e.g., Opus-class) are becoming 'hungry,' token-heavy, and increasingly premium-priced, and will remain valuable for open-ended generative work. But for repeatable business processes, firms will likely need more specialized, utility-class local models (comparisons made to Llama and DeepSeek) that require more technical expertise to deploy as agents that behave consistently every time - which is described as very hard to achieve with generative models because consistency fights against their generative nature. The current approach described is using AI for two things: turning messy real-world data into a structured database, and then calling that database to perform actions - deliberately avoiding letting AI do any math itself, relying instead on database routines. Owning your own models/servers/compute was described as a likely future step, but not economical yet since compute is still cheap. The prediction is a market split: specialized local models for domains like accounting or law, versus premium generative models for open-ended tasks - effectively pushing firms toward building their own customized 'apps' around AI where domain expertise and data become the differentiator.