Natural language processing (NLP) may indeed understand the unstructured text, then according to (2), the "Semantic Web" is not needed, except for perhaps cachi

Natural language processing (NLP) may indeed understand the unstructured text, then according to (2), the "Semantic Web" is not needed, except for perhaps caching NLP outputs in machine-readable form. (1) is more fundamental: a lot of value-add annotation (in RDF or other forms) would be valuable, but because there is work involved those that have it don't give it away for free. This part was not sufficiently addressed in the OP: the Incentive Problem. Either there needs to be a way how people pay for the value-add metadata, or there has to be another benefit for the provider why they would give it away. Most technical articles focus on the format, or on some specific ontologies (typically without an application). A third issue is trust. In Berners-Lee's original paper, trust is shown as an extra box, suggesting it is a component. That's a grave misunderstanding: trust is a property of the whole system/ecosystem; you can't just take a prototype and say "now let's add a trust module to it!" In the absence of trust guarantees, who ensures that the metadata that does exist is correct? It may just be spam (annotation spam may be the counterpart of Web spam in the unstructured world). No Semantic Web until the Incentive Problem and the Trust Problem are solved.