Picture, if you will, a lumber company that sells its own products online. Like the CPU manufacturer, they grade their lumber into categories including FAS (the best), No.1 (furniture grade), and No. 3 (flooring). Obviously FAS is the most expensive, but they don’t sell a lot of it. Since the goal, when milling logs, is to get the most and best woods from them, they often have a surplus of a FAS and run out of Nos. 1 and 3.
If they list the Nos. 1 and 3 as out of stock, people with lesser needs and smaller budgets won’t step up to the FAS, but go elsewhere. The store could offer the FAS at the same price as the out of stock wood, but then they would lose the money from those ready to buy the more expensive product at the original price. So the store starts putting better grades of lumber, at random, in with the lower grades. That way they can sell as much wood for as high a price as they think possible, which is just good business.
If a customer notices that some of the wood is better than what they paid for, they are unlikely to complain. In fact, the lumber firm could develop a reputation for selling better than expected products.
Making sure we get all the saw dust off before we go from the mill back to the semiconductor fabrication cleanroom, we now see that our CPU could indeed be FAS when we only paid for No.1, or a 3.3+ GHz when we only forked out for a 2.8 one.
Overclocking, at least while maintaining stock voltages (more on that later), is not really about pushing the envelope, but discovering the chip’s actual potential.