Watson is a red herring.
Imagine if Watson was an app. What would you do with it? As I understand it, it takes an oblique textual reference of some kind and turns it into a concrete, explicit reference. You talk about something and it figures out what you're talking about. That's great! It's a hard problem and it's extremely impressive that it's been solved.
But in practical value terms, that's just a toy. And when would it be more useful than Google Search, or Wikipedia, or Siri? Can Watson do anything, book a trip, set a timer? Well, that capability would have to be plugged in, along with voice recognition and speech synthesis. But can it reliably determine what I want it to do? Or even whether I want it to do anything? Can it follow a conversation? Is it even stateful? What, if any, role does the core Watson functionality play in the systems which eventually do these things?
And how does Watson generalise to other areas? IBM was looking at applying it to healthcare. How would this work? Presumably, here your "oblique textual reference" is a collection of data (symptoms and measurements) collected from a patient, and the end result is a diagnosis, like "What is rubella?". Now, firstly, how useful is that? Is that the hard part of being a doctor? What about carrying out the tests, what about prescribing treatment or choosing further tests? What about missing hints, e.g. can Watson handle visual data from simply looking at a patient? Or does the doctor have to inform Watson that the patient is leaning slightly to her right when she sits, and has trouble getting onto the bed? So does Watson's single trick actually help, does it even slightly alleviate the workload of a medical professional?
And secondly, how readily does Watson adapt from the world of general knowledge quiz answers to the world of medical diagnoses? Are these worlds even remotely structurally similar? Can they be modelled and correlated using the same basic structures? Is there some heavy modification and serious domain knowledge needed to alter Watson to do this? Or are the two problems basically on different planets, such that an entirely different machine needs to be built?
The same goes for applying Watson to business analytics, for example. In my mind, Watson is a machine which, as it currently exists, best case scenario, can look at all of your financial graphs and go "Ding! You're in a recession." And then nothing else. This may be of some use, but I doubt it.
All of this is ignoring the backend, which is that Watson in reality is this monumental pile of expensive, high-performance hardware and IBM Research code - and we all know how much research code ever resembles a working, saleable product. All of that to handle one real-time quiz game. You're going to shrink that until it fits in one office? Or you're going to have a server farm full of them serving requests at great expense?
I feel like people perceive Watson as a hair's breadth from strong AI, and IBM doesn't want to disabuse anybody of that notion. But I think the really valuable stuff in Watson - advances in computing techniques which I don't really know anything about but I'm certain must be there - is far less tangible, and is going to be very difficult to extricate from Watson itself, let alone monetize.
Just off the top of my head. Unsourced hunches.