We can't have that...
The bottleneck is the human effort to deploy the tokens.
I imagine that bottleneck is much, much, less of a constraint for openAI, anthropic, and the ilk.
"Put the software factory and reverse engineering machine in the rack bro."
DeepSeek just open-sourced its inference engine - https://news.ycombinator.com/item?id=50163277 - October 2026
Also, how much would be typical AI token costs of doing this for various apps, simple to complex using some reference AI model?
Thanks.