What the reasons are is largely a matter of opinion. In my opinion there are several possible reasons: - the 'semantic' idea of 'strong' modeling of the world lost out to a competing approach that uses probabilistic models. The latter models don't require coordinated effort by humans and thus scale better. - the semantic approach also suffered from academicians myopically going over the same millimeters of theory for decades. And losing sight of matters of practicality. - The fundamentals of the semantic technology seem rather brittle to me. With that I mean that a tiny difference in the reasoning axioms can make the whole reasoning intractable. That might be anti-tethical to the 'tinkering until it works' approach that software engineers often use. - Somehow a broady applicable killer application didn't turn up. But that's a result as much as it might be a reason.
There's still use for the technology though. If you need to unify data that follows subtly different datamodels I'd be hard pressed to think of an alternative. Which makes me wonder whether intelligence agencies use the technology. F.i. I remember a reference given by Oracle of the US Geospatial Intelligence Agency using their quad store. The new moniker 'linked data' emphasises this aspect. Government agencies do struggle with having to relate data that are obviously related but conceptually, legally subtly different. And they do spend quite some attention on linked data. They seem to stay mostly within the RDFS realm and do not stray into more interesting OWL applications. But even in government circles I get a whiff of the solution-looking-for-a-problem vibe that hounded the semantic web for so long.