By my two-second calculations, if about half of the worlds population were using Bitcoin and making 2 transactions per day, a flat average would be about 70,000 transactions per second. To account for peaks which one typically finds in network service use, better bump that up to about 250,000 transactions/sec. Then one must consider the potential for DOS attacks.
Contrast this with the .25 (aka 1/4) transactions/sec at the current high.
As currently implemented, Bitcoin scales until it fails then that's all she wrote. Ultimately I would be more comfortable with a solution which anticipated 'sharding' of some sort, and it seems logical and workable to combine this with a multitude of crypto-currencies each tuned to offer various advantages to their user base.
Of course the 'money changers' would have a hey-day in such an environment, but that fine. Probably open systems would develop to make things less costly for end-users. Some guy had a link to such a project a while ago. I think it was on Docmeister's Bitcoin-2 post.
Vast majority of those 250000 transactions are going to be small transactions that need instant confirmation, which means they will not be handled directly by bitcoin network. There will be a bunch of payment processors who will provide instant confirmation functionality, and use bitcoin network itself to regularly balance the books between themselves.