Every line of code or processor instruction cost. The more efficient algorithm, the less operations (or 'algorithmic steps') are needed to achieve certain goal. Cost is measured by function, for example linear or linear-algorithmic or faster or much faster rising... that is, the more data to handle, the more it matters.
For example... we want to calculate function of 120 variables. Our algorithmic cost is: n2... the total cost of computing this result is 14400 steps. If algorithmic cost was for example: (3 / 2) * n, the cost of computing such function would be 180. With larger values, algorithmic cost matters much more than buying more computers and their speeds.
Customers or system designers can decide where to put these processor cycles and memory... into pretty graphics, security (for example, securing broadband connections with encryption on-the-fly) or something else. However, half secure system is much easier to break according to experts, than well done one... so it's best to focus on whole security in all aspects than do it half way. Cost-wise it's also more efficient.
For more see Literature , , perhaps more.