Most of us remember from childhood Robert Southey’s tale, “The Three Bears,” where Goldilocks searched for the perfect bowl of porridge, chair, and bed in the house of the three bears, testing each until she found the one that was “just right!” I often think about this tale when considering data center infrastructure planning. Infrastructure planning is the careful art of ensuring a balance between available infrastructure supply and customer demand, while optimizing operational and capital costs, using the Goldilocks Principle.

Traditionally, the focus of data center operators was on operational uptime levels and management to meet uptime availability targets. However, as the scale of cloud infrastructure reaches global proportions and availability moves in the application layer, the focus has turned to efficient use of capital in delivery of cloud infrastructure and services. Taking a page from manufacturing and retail industries, which optimize product delivery across the entire supply chain — from manufacturing and distribution to retail delivery — how can we apply supply chain concepts and methodologies to the cloud industry in order to optimize delivery to data center customers?

In the realm of data center (DC) infrastructure, there has historically been a heavy focus on maintaining uptime, while optimizing power and cooling, and driving towards progressively lower power usage effectiveness (PUE). However, from the perspective of supply chain, a service level of 95% indicates that 95% of the time customer demand will be serviced with available stock. For example, in the world of retail, if a store guaranteed a 95% service level for a particular product it would mean that when a customer selects the product (in store or online), the product is available to promise 95% of the time. Clearly, a higher service level requires a higher amount of inventory, or a responsive supply chain that can quickly deliver product to the consumer. From this perspective, if a consumer of cloud computing, storage, or physical servers expects a service level of 95% or higher, a supply chain approach to capacity delivery is recommended.

This article will discuss application of a multi-echelon supply planning model, then discuss the data and supporting business processes and policies required to support the model. Lastly, I will provide some guidance on how to begin implementation of this approach, as well as challenges that are common. This is a brief adventure into multi-echelon planning and optimization, and attempts merely to start the discussion, and suggest possible approaches. Each business will need to develop a specific approach using some of the guidelines provided.


Multi-echelon planning is a supply chain approach with the goal of balancing inventory costs across an entire supply chain. For reference, articles abound on application of multi-echelon planning in retail distribution networks and manufacturing.

In this approach, the supply chain is divided into levels, and within each level of the supply chain, inventory is optimized through application of analytics to drive order quantities, maintain minimum and maximum inventory levels, and trigger additional orders when inventory falls to minimum levels. As an example, a particular regional distribution center will maintain target inventory at lower levels due to lower regional product demand as compared to other centers.

When planning overall data center infrastructure, each firm may have a different definition of the echelons or levels within their supply chain based upon the business model of the firm. For illustration, here are a couple different examples tuned to generic types of data center businesses:

Example 1. Infrastructure-as-a-Service (IaaS) provider, selling pre-built physical server capacity, volume at ~100 servers per month across four U.S. regions. Business model is to lease capacity from data center supplier, prebuild racks, network, and commodity physical servers of a variety of types.

In this approach, there are four levels of inventory to be optimized: leased cages, racks, network, and physical servers. The firm analyzes the time to deliver new leased cages, new racks, network build, and physical servers, including standard deviation, as well as the overall demand for new physical servers. Based upon the analytics, orders (AKA builds of additional leased cages) are triggered when the months of supply of empty leased cages hits the minimum level. Similarly, additional racks are ordered and deployed when empty rack inventory reaches the minimum level. Lastly, servers are ordered when the days of server supply hit the minimum level. Ports on the network are similarly monitored, and additional blades or switches deployed when needed. Each region has different demand trends, therefore, the amount of supply planned in each region is aligned to match the demand requirements. By taking this approach, capital expenditures are optimized across the supply chain, and inventory carrying costs are minimized.

Example 2. Wholesale data center provider. Historically, many of these companies have built new data centers in a standard size, regardless of location. The multi-echelon approach depends upon the ability to build data centers in a modular fashion, sizing the building to address the demand in a particular location. Each module size would need to be designed for efficient power and cooling, and the multi-echelon planning approach will govern the timing of each additional modular build. Depending upon the demand for each location, the number of modules to be built is governed by the order quantity modeling and min-max inventory levels.

In this particular instance, there are three echelons proposed in the supply planning model. Again, each echelon is independently monitored and economic order quantity (EOQ) modeling, minimum and maximum supply levels, are maintained. The end result is optimized inventory of land, shell, and finished colocation rooms depending upon the demand and supply characteristics for each location. Varying modular build sizes are launched depending upon the demand for a location, as well as the remaining months of supply. A particular piece of land may hold multiple powered shells, and depending upon the time required to perform site selection activities, land supply levels will be monitored and new site selection ensure when required. Taking this approach again optimizes capital expenditures and carrying costs, and aligns the supply plan with demand variability by location while ensuring customer needs are met.


The basics of implementing multi-echelon inventory optimization to DC infrastructure planning are standard supply chain analytical equations to determine order quantities and safety stock levels, based upon variable demand and lead-times. In order to develop the analytical models, clearly there is a heavy dependency upon data.

Standard EOQ models are built using a standard set of data inputs, which would need to be available for each of the levels within a multi-echelon supply chain. Table 1 lists the standard inputs with a description of the data required to build the model for a new powered shell.

Once all data is collected, the model can be built, and solved for the ideal order quantity, expressed as Q*= optimal order quantity, which in the case of powered shell, would indicate the ideal build size for a given location, depending upon the data inputs for the specific location. Of course, depending upon the planned generator sizing, the ideal order quantity will need to be compared to the nominal generator output, and “rounded” against the generator size planned for the site. The basic EOQ formula, solving for the ideal order quantity, is found in Figure 3.

However, demand is typically variable, and also supply lead-times (e.g., the timeline to build a new powered shell) can be variable depending upon many factors in construction or infrastructure build-out within a particular site. In this case, more advanced models can be developed to determine the build sizes, timing of next build, and amount of supply to hold as safety stock. Looking at the levels proposed in Example 2, data will need to be collected for each level, to provide the inputs to advanced modeling, and further optimize inventory across the infrastructure supply chain. For the demand variability, looking at total allocated (fill rate) capacity of a given inventory echelon vs. customer forecasts and internal forecasts will mean collection and storage of data from DCIM, CRM, and other systems. Similarly, the supply lead-time, project, and workflow data will need to be stored and collected in order to further optimize the inventory models.


There are many different business models in the data center industry, from firms that sell powered shell, to IaaS providers, cloud service providers, colocation providers, and a variety of others. Each firm may hold a complex mix of current inventory, planned future inventory, and also plans to enter new markets. Depending upon the age of the firm, a preponderance of legacy inventory may also be owned. Determining the levels within the infrastructure supply chain for inventory planning also assists in sorting out longer-term plans in terms of when to decommission or refresh particular infrastructure, and by providing a framework for better understanding of the overall complexity of infrastructure held vs. the specific demand requirements of each customer.

In order to determine the appropriate number of echelons, the infrastructure being delivered can be sorted using the following comparison matrix: Cost per unit, lead-time to deliver, and flexibility to serve customer demand.

Doing an analysis to sort through which infrastructure to include in each level should not be too complex. The primary goal is to delay the most expensive, shortest lead time, and most customized infrastructure to achieve more efficient use of capital, which in the chart above would mean holding land until a new powered shell is required, and then delaying ordering network equipment until closer to the time that customer demand must be fulfilled. Obviously, most businesses are already making these choices, but applying a systematic approach backed up by data and analytics leads to improved business results.

It may seem that implementation of multi-echelon planning would take years to design; however, I always recommend getting started with the data available today, and then building upon this in phases. Put together a simple list of the infrastructure building blocks, and determine whether demand, lead-time, and cost data is available for each set of infrastructure. If you have all the data today, you are a lucky soul, and can quickly begin defining the echelons in your supply chain and building basic analytical models. If not, and data is only available for one type of infrastructure, then simply start building analytics around this data, and define a plan to begin collecting data across other infrastructure types as soon as possible.

In summary, I hope that this article gives some insight into how Goldilocks would plan for the “just right” size of land, rack rows, powered shell, or number of physical servers to order and hold within an infrastructure supply chain. Each company is unique, and needs to consider application of any new approach based upon business needs.