The problem is the way consumer rights are going, companies will simply make it impossible for you to run a line like that, or at least make it difficult enough that only a small percentage of the population can do it.
Every layer of friction they add means millions of people just throwing up their hands and accepting whatever nonsense is thrown at them.
You assume all models will be closed-source and there will be no competition. I can imagine that, but I would assume training an AI model for ad-filtering is something the community can and will do - maybe think 5 years ahead what hardware will be available for training and inference to a "small developer".