There's other factors too. Example: if you don't need fast but you have a specific domain you're working in, you can throw regular RAM and CPU inference at doing things with a defined pattern.
For example, no software engineering manager wants to throw hours of work at refactoring declarative infrastructure-as-code to modern tooling from old shit - it's hours of work to arrive at the same place, but with better capability and supportability; classic tech-debt everyone loves to ignore until it's too late. So, instead, create a queue and a few mid-grade servers behind it running qwen3-coder and give it a good tuned prompt to convert some ancient CloudFormation shit to modern Pulumi in Go, using shared modules that you already wrote to abstract complexity and cloud best practices / secure design. Who gives a shit if it takes 3 hours to figure it out - that's still far less time than it would take an actual engineer to do it, and it's probably a best-case scenario for an AI: a tight declarative 1:1 target where you can compare the previous stack to the `pulumi preview` and know if you've hit the target or not.
Your engineers get to spend time doing useful productive product development where their time produces tangible value, and you get policy-compliant imperative IaC code for the cost of the electricity to run a few boxes you already own blinking away in the night, reducing technical debt.