I see two areas of application. One is a place for startups and lifestyle businesses, the other is a place for government, big businesses and consultants to them.
(1) Using data that's already available in RDF or RDF-like format.
The most important examples in my mind are DBpedia and Freebase. There are hundreds of fun and informative web sites to be created based on those data sets, although you'll need to invest in data cleaning to make anything commercially viable.
Generic databases can also provide world knowledge to NLP systems -- DBpedia Spotlight, for instance, is an NLP named entity recognized system based on world knowledge that's (almost) entirely ignorant of grammar. After a year of development it's competitive with commercial grammar-based systems that have been under development for decades... And I believe it's going to make more progress in the next two years than conventional systems.
Unfortunately, Facebook's Open Graph protocol/API isn't open enough to enable interesting applications. For instance, Facebook has created pages for most topics in Wikipedia -- these can, on an individual basis, be reconciled to DBpedia topics, but there's the complication that often the "official" page for something in Freebase is a page managed by the entity, not the one from Wikipedia. With a reconciliation API or NT file, it would be easy to crosswalk Facebook pages with Dbpedia and Freebase data, but as it is, it's not easy.
(2) "Enterprise" Data Integration
All the time there's some story in the news about how a big organization is doing a project that involves integrating a large number of legacy systems. For instance, CALPERS (California's State Pension System) is integrating more than 60 legacy systems to make a system that will be helpful to state employees as well as CALPERS workers. Projects like this always run over budget and over schedule, and it would be great to have something that lowers cost and squeezes out risks.
I think a system based on declarative and logic programming could be a big improvement over the status quo. RDF is a good candidate for a 'universal' data model that can express information that exists in different kinds of databases.
One mode is to deal with a limited amount of information at a time (say a single customer) so that we don't abuse the scalability limits of highly expressive reasoners. The Amdocs example that Craig brings up is a good example.
Another mode is to 'throw it all into a triple store.' Many data warehousing applications use Star Schemas and bitmap indices which are similar in organization and performance to many triple stores -- RDF technology can be more competitive in this space than it is in OLTP.