Let's say it's 20 years from now, the total size of the bitcoin economy is $1 trillion, and there are 10 million bitcoins in existence. (We could find the number of bitcoins exactly, but 10 million makes the math easy and it's within a factor of 2 of the ultimate limit. There are about 6 million coins right now.)
$1 trillion / 10 million coins = $100K per coin, so we can expect people to spend almost that much money to generate a coin.
The rate of coin production is currently 50 per 10 minutes, dropping in half every four years, giving us 9 coins per hour in 2031.
That's 78K coins per year, at $100K per coin, or $7.8 billion in computer time and energy spent per year.
Let's say half of that is energy, call it $4 billion, including the energy to make the dedicated mining computers. At ten cents per kilowatt-hour, we're talking 40 billion kilowatt-hours per year. Divide by the number of hours in the year and we get 4.5 gigawatts.
In other words, a $1 trillion bitcoin economy can be run on the output of roughly five typical nuclear power plants (assuming it's 20 years from now).
If we get to $1 trillion in only 12 years, energy usage will be four times higher, since coins will be worth about the same but four times as many will be generated per hour.
If people use custom ASICs to generate coins, they'll be costly but more energy-efficient, weighting expenditures more heavily on hardware than energy.
People can grant transaction fees to miners to make their transaction go through faster. If transaction fees become prevalent, resource usage will be higher, since there will be additional reward for completing blocks. But $7.8 billion is already almost one percent of the bitcoin economy. If the average bitcoin changes hands once per year with a one percent fee, or ten times per year with a 0.1 percent fee, we add $10 billion to mining rewards and approximately double our resource usage.
Edit: $1 trillion is about equal to the amount of U.S. currency in circulation: http://en.wikipedia.org/wiki/Money_supply