Ad exchanges take ad impressions and let many potential buyers place a bid for said impressions. The buyer who bids the most wins the impression.
Right now, some ad exchange inventory (i.e. Right Media's Yahoo inventory) auctions off only non-guaranteed inventory. So if you step up and pay rate card or near rate card prices, you are guaranteed that inventory. If not, the inventory goes into an exchange, and the highest bidder gets it.
Participants in this exchange for non-guaranteed inventory include not only real-time bidders, but networks and other companies that have more fixed-price type contracts with various advertisers. One advertiser may be willing to pay $.75 CPM for an above-the-fold leaderboard on a top 1,000 sports site with a frequency cap of 2 per day. So this is entered as a $.75 (or less) bid for every impression that fulfills the above requirements.
As the market matures, though, you'll find that exchanges won't only include "non-guaranteed" impressions in their exchange marketplace, they will essentially include "guaranteed" campaigns in the exchange and just make sure they charge enough such that these advertisers win the auction enough of the time, essentially "guaranteeing" delivery. Good forecasting and data will enable them to make set these market prices for "guaranteed" inventory. If your rate card is $5/CPM and an ad exchange participant can pay $6/CPM for some impressions, why not sell it to the higher bidder (or raise the price as appropriate to the guaranteed buyer)?
So where does this leave the unsophisticated buyer, buying on more traditional contract terms, in essentially a non-real-time, or pseudo-real-time through an intermediary basis? You'll end up with impressions that meet your contract's criteria, but have very little value. Depending upon what criteria you fail to specify (and you can never specify them all), you'll get high frequency users (with notoriously low CTRs), younger users (even if the exchange doesn't know their age, someone else will), placements that are below the fold and go un-noticed by users, and impressions from users whose value is less then average due solely to the fact that no other advertiser has found an attribute that would make them look valuable. Even if there's no known attribute that makes an impression less then average valuable, the fact that there's no attribute that makes it more then average valuable (known by any participant bidding on the exchange) probably means it's worth little.
In summary, if you buy display advertising, get sophisticated, and get sophisticated fast. Understand exactly what user actions or conversions you associate value with and make sure you or a company on your behalf can correlate this value to every single data point that you or they or the data exchange has on a possible impressions. Those who fail to de-average groups of inventory will be stuck buying the slop, and firms who properly value inventory across a huge number of dimensions (both declared and proprietary data dimensions) will benefit significantly in this change in the way advertising impressions are allocated to buyers.
And just you wait until the search platforms let people real-time-bid on a much larger set of variables then are available to search advertisers today. How many relevant mathematicians are on your team, that can handle machine-learning and statistical inference type problems, utilizing data from advanced data mining?
Use ad exchanges to effectively de-average the value of impressions across a huge number of impression attributes (and bid appropriately), or get stuck with what no one else wants...and watch your ROI and volume get whiped out.