I worked on the Semantic Web, designing core data infrastructure, back when it was still hot. It disappeared because it had two fatal flaws: it was intrinsically non-scalable, both conceptually and technically.

First, there is no universal "semantic". The meaning of things is ambiguous, and given a broad enough pool of people and cultural contexts, it becomes nigh impossible to converge on a single consistent model for individual terms and concepts in practice. A weak form of this is very evident in global data interchange systems and implementations, and the Semantic Web took this bug factory and dialed it up to eleven. In short, the idea of the Semantic Web requires semantics to be axiomatic and in the real world they are inductive and deeply contextual. (Hence the data model rule of "store the physics, not the interpretation of the physics.") That said, it is often possible to build an adequate semantic model in sufficiently narrow domains -- you just can't generalize it to everything and everybody.

Second, implementation at scale requires an extremely large graph database, and graph database architectures tend to be extremely slow and non-scalable. They were back then and still are today. This is actually what killed the Semantic Web companies -- their systems became unusable at 10-100B edges but it was clear that you needed to have semantic graphs in the many trillions of edges before the idea even started to become interesting. Without an appropriate data infrastructure technology, the Semantic Web was just a nice idea. Organizations using semantic models today carefully restrict the models to keep the number of edges small enough that performance will be reasonable on the platforms available.

The Semantic Web disappeared because it is an AI-Complete problem in the abstract. This was not well understood by its proponents and the systems they designed to implement it were very, very far from AI-Complete.