crawling the web to build the database is trivial
Well, trivial in the sense that it doesn't require particularly sophisticated technology. Just huge, gigantic piles of not-sophisticated technology. The deployment, provisioning, and maintenance of that huge pile of technology is non-trivial. The code itself is relatively trivial, though. But code is only a tiny portion of what's required.
scaling watson up to hundreds of millions of users is the problem
Exactly. And this is even more so when you consider ongoing costs. Google uses cheap, throwaway servers in mind-boggling quantity. Building gigantic clusters of Watson-style machines is going to be an entirely different proposition. Between the up-front costs and the on-going maintenance costs, IBM would need some serious motivation to even think about trying this. Frankly, it doesn't seem like it fits their current business model at all.
I might be able to see IBM trying to license Watson-like technologies to Google, MS, Yahoo, and/or Ask. I can't see them suddenly switching direction to the degree that would be required for them to try to compete directly.