Watson on CUDA? Could a smaller version of Watson be written for CUDA

I watched on the edge of my seat as Watson Stomped a mud hole in the jeopardy champs then walked it dry, I loved it and I started to think with the massive Parallel processing Power of CUDA could a Watson Like program be written?
I don’t want to play Jeopardy but it would be very nice to be able to talk to a computer and have it understand the nuisances of English language and even understand a joke. IBM did a great job those people are just scary smart my hat is off to them but could an open source program be written to one up Watson? With all the twist and multiple meanings of works in the English language and taking the work context into the equation.

IE. What is proper to say 5 and 6 are 12 or 5 and 6 is 12 and the answer of course is neither because 5 and 6 is 11.
A Lite Version could use the internet as a database and sure it wouldn’t be able to answer every question in less than ten seconds but it could answer them. If I ask my computer what is the name of that song that goes do dod dee do do do and repeats the world banana over and over it should be able to understand my question and find the answer.

Are you sure the bottlenecks are in the computation and not in for example internet bandwidth or disk read bandwidth ?

Markov chains should fit nicely for speech recognition and i believe these are very vector like in nature.