Instead, there are many decentralized hubs called routers that direct data from one local network on the Internet to another.

Each piece of data sent through the Internet has a number to specify one of four billion destinations. Routers refer to a road map--called a "next hop table" in computer terminology--to point the data in the right direction.

When one Harvard students sends another an e-mail message, Harvard's routers know to move it from one Harvard computer to another. This data never leaves Harvard's section of the Internet.

But if a student wants to send an e-mail message outside of Harvard, the university's routers cannot handle this request alone. They send it to a router outside the university to see if that machine knows where to send it.

Today, the next hop tables in the memory of central routers have 90,000 entries. This may sound like a small number with tens of millions of computers on the Internet. But creating a next hop table with this many entries is a computationally intensive process. And each time a new listing is added to the central next hop tables, the entire table has to be recalculated.

The time and computing power it takes to recalculate the central next hop tables after each additional new entry increases as the tables get larger.

The problem, Bradner says, is that the Internet is growing in size exponentially. He argues that in about two years it will take so long to recompute the Internet's central next hop tables that it will be impossible to expand them.

"You need to increase computing power more than is practical and faster than you did earlier on," he says. "At some point you will just have to stop adding new networks."