Hypothesizing: it's doing things which require decidedly non-BigTable-like architecture. Most other Google apps in their alpha stages can just take advantage of running on, basically, App Engine; this seems like one of the ones that can't, so it has to run on a separate architecture that hasn't been scaled yet. That would be alright for some other company's demo—if they were anyone else, they'd just use limited, cached, pre-defined data sets. But they're Google, they have the raw data, and they wanted to show off the possibilities of working "from scratch" on anything they like in realtime. Thus, it was similar to an app without any scaling consideration as-of-yet (consider a normal startup webapp) drinking from the entire search index firehose on each query. Ten seconds is impressive.

Once it's deployed across Google's entire cloud it'll speed up nicely.