My guess is that they predict the most common searches and constantly query and cash results for these searches, something I don't think Kayak or hipmunk does.
I don't think you can launch a search product on google if it takes more than .xx seconds to return the result...
I'm not so sure, I looked for a few random flights. However, having built my own very rudimentary internal flight search tool from the global flight list, I can tell you there are roughly 7 million flights per month, a small data set by google standards.
It's not the size of the dataset that is relevant but more that this dataset changes in realtime : price + availability. These requests go through GDSs that are relatively slow, that's why so called real-time search engines like kayak are fairly slow. What do you exactly mean by 7M flights? Keep in mind that most search requests involve a combination of flights (leg), outbound, inbound and sometime multiple stops...
I think that owning ITA allows them to have more data to mine/predict and better access but it doesn't explain how fast the searches are.
Most airfare searches still hit GDSs which are relatively slow to respond. The only way I can think of Google can deliver such a performance is by intelligently pre-caching airfare searches. Google?
And be sure to click around. It has some interesting tools, like prices for flights to your destination flown out of nearby airports.