The Value of on-Demand Services
One other key attribute of the cloud as we’ve defined it is on-demand resources or services based on those resources. Briefly, the value of “on-demand” arises from avoiding both excess resources and insufficient resources (Weinman, 2011f, Islam, S. et al., 2011). Excess resources are costly due to the weighted average cost of capital used to acquire the resources, or the opportunity cost of the capital not being productively employed elsewhere. Then there is the risk of obsolescence, requiring premature write-offs, the risk of loss, or the cost to ensure those resources against loss. They require floor space, and often require power and cooling.
Conversely, insufficient resources mean lost revenue, poor customer experience, loss of brand equity due to poor customer experiences, and in defense applications, inability to support the mission. These costs may have nonlinearities.
Let us briefly consider symmetric linear penalty costs associated with insufficient and excess resources. In effect, if the demand is D and the resources are R, the penalty cost is proportional to to | D – R |. If they are time varying, the penalty cost is proportional to 
This will make analysis more tractable.
If demand is flat, on-demand resources are unnecessary.
If demand grows linearly and predictably, even long provisioning intervals are acceptable: one merely need order additional resources on a regular schedule.
Conversely, if demand declines linearly and predictably, if resources can be deprovisioned without extra costs (from packing, restocking, auction, etc.), on-demand is not necessary.
It is when demand is unpredictable and/or non-linear that issues arise with traditional resourcing and on-demand really shines. For example, let demand in a given time period be uniformly distributed from say, 0 to 1. If resources are set to 0, the penalty will always be due to insufficient resources and the expected value of the difference will be ¹⁄3. If resources are set to 1, the penalty will always be due to excess resources and again, the expected value of the difference will be ¹⁄3. However, if the resources are set at the midpoint, the expected value of the penalty will only be ¹⁄6 = ½ × ¹⁄3 × ½ + ½ × ¹⁄3 × ½.Of course, if the right resources are available at the right time, the penalty is zero. If we scale this model up to a uniform distribution on [0, P], and assess the penalty over time T with a penalty cost of c, it is clear that the total penalty for fixed resources, even if we can set the level optimally, is ¹⁄6P × T × c. If the penalty function is linear but asymmetric, for example, the cost of unmet demand is higher than the cost of unused resources, the optimal fixed strategy tilts in favor of a “better safe than sorry” approach of excess resources, with an optimal point being distribution-dependent.
Space does not permit detailing the calculations I’ve done elsewhere (Weinman, 2011f), but we will highlight conclusions. When the demand function is exponential, i.e., D(t) = et, any fixed provisioning interval that attempts to provision in accordance with the current demand level (i.e., there is no forecasting) will fall exponentially further behind. This is because if the fixed provisioning interval is k we set resources R(t) = et–k, and thus the difference is D(t) – R(t) = et – et–k = et(1 – e-k) = k1et, where k1 = 1-e–k, and thus the penalty cost is c × k1et, in other words, grows exponentially over time as well.
The Value of Online Connectivity
Connectivity is an enabler of the prior value drivers. After all, on-demand resources can’t be shared unless various customers in various locations can access them. Without sharing, pay-per-use pricing is economically unattractive for service providers to offer.
Connectivity then, has values and costs. When examining connectivity the costs may be clear: dollars per Gigabyte transferred or the capital costs of routers or optical facilities. The value is harder to quantify, since it often is an externality. A good approach is to consider the marginal cost, if any, of connectivity, and use it to offset the benefits associated with specific other pure cloud or hybrid cloud approaches. Often, this will be easy: an ecommerce spike can be probability-adjusted, the revenue associated with it estimated, and the marginal benefits associated with partial or total cloud solutions thus quantified. Against this must be deducted any marginal network costs to assess the value of alternate approaches.
Behavioral Cloudonomics
We will now turn to some ancillary considerations in cloud economics. A basic assumption above is that neoclassical economics, expected utility theory, and quantitative finance rule decision-making. If, say, the expected, i.e., probability-adjusted, net present value of choice A costs less than that of choice B, we should select it.
A more recent approach to economic decision making is behavioral economics (Ariely, 2008, Lehrer, 2010) which admits that humans do not always make purely rational and quantitative decisions, but that often various cognitive biases and heuristics play a role, as well as “bounded rationality,” the limits of people to calculate appropriate answers to decisions. In this field, numerous experiments have been conducted, books and scholarly papers written, and Nobel prizes awarded. We can’t do the topic justice here.
However, we note that a number of behavioral economic favors impact the cloud—what may be called Behavioral Cloudonomics (Weinman, 2010). For example, perhaps the most well known result is that of “loss aversion:” people generally get less satisfaction from gaining a dollar than they feel pain from losing one. As a result of this and other factors, customers may recognize the financial advantage of pay-per-use, but avoid it due to what has been called a “flat-rate” bias (Lambrecht and Skiera, 2006). For example, the fear of an unexpected large monthly cell phone bill can lead one to overpay for services by signing up for a flat-rate plan even though measured service—i.e., utility pricing—would be cheaper due to normal light usage. Such biases could negatively impact cloud adoption.
On the other hand, they can also work to the cloud’s advantage. Dan Ariely and his colleagues at MIT have conducted experiments regarding the special attraction of “free” (Shampaner and Ariely). Again, rational behavior is subsumed by irrational: subjects typically select a free ten dollar gift certificate (a ten dollar value) over paying seven dollars for a twenty dollar gift certificate (a thirteen dollar value). The lack of upfront investment in using the cloud is thus extremely attractive as it aligns with this particular bias.