I've talked to a number of founders about wanting to launch a $TAO subnet. I've identified a rubric (VCs, amiright?) for what would make for a successful subnet based on taostats.
Some subnets are clearly more productive than others and well suited for Bittensor.

ethereum protocol research /ethrd + Crypto x AI /dacc 

Bingo! And this isn’t because general models or LLMs can’t be served. The fast feedback nature of other models will drive emissions towards those high throughput subnets.

So no generative models & LLMs basically. From what you write, traditional ML applications such as forecasting/classification seem better fits.