235 Montgomery, Suite 1017 San Francisco, CA 94104 (415) 373-0673 www.ramprate.com The 7 Principles of Data Center Cost Metrics The Cost Evaluation Methodology Vendors Wish You Didn t Know By Alex Veytsel Principle Analyst RampRate August 2010
When planning a data center outsourcing strategy, many elements need to be taken into account, including geography, data center tier, historical reliability, reference reviews, SLA guarantees, protection against natural disasters, etc. A best practice approach boils all of these considerations down into a simple scorecard for direct apples-to-apples comparisons across all aspects of potential solutions, including both outsourced and in-house. While a full evaluation is a complex process requiring years of market data, vendors historical performance, and highly technical sourcing experience, one basic facet of the selection process can still cause trouble even for some of the largest and most experienced colocation buyers: identifying a common size metric and cost accounting model. This paper addresses the methodology and reasons for looking at cost of data center / co-location space per peak kilowatt as the preferred method of cost accounting, benchmarking and evaluation of data center fixed costs (including rent, maintenance, and depreciation of non-it equipment such as UPS and generators). This paper concludes with a Q&A that we compiled from the common questions we received from customers seeking to adopt data center outsourcing costs evaluation best practices. Fundamental Goals of Optimal Cost Metric in Procurement In order to say which analysis model is better, we must first ask "better at what?" We ve learned that a cost analysis metric should follow 7 key principles: This is essential to enabling comparisons between different vendors and facilities purely on a financial basis, but is often very difficult to achieve due to confusing pricing models, opaque cost drivers, and conflicting reporting parameters. 1. Comparability of quotes The metric should allow apples-to-apples comparisons, so that an offer from one provider, location, or project can be contrasted with offers from other providers in other locations and projects. 2
2. Flexibility to spec changes The metric should allow for inferences about the impact of design changes on costs without going back to vendor. 3. Correlation with underlying cost drivers/constraints vendors should not be able to increase chargeable amounts to the customer unless they are incurring additional underlying costs. 4. Vendor neutrality Proprietary terminology from one vendor precludes cross-vendor comparisons, so the metric should be neutral to all vendors or options. Potential Cost Models for Data Center Space The underlying elements in data center space costs are always the same land, facility, equipment, maintenance, power generation/consumption, and overhead. However, depending on the vendor's background, you may see one or more of the following six price models for this same bundle: 1. Per deployment -- a total bottom line quote for a customer s specific data center spec (e.g. the total for X racks with Y power is $Z per month) 5. Translatability The metric should be clearly definable and translatable into other charge models. 2. Per rack (or cabinet, or rack unit) a. 7 tall b. 10 foot tall 6. Minimization of misunderstandings The metric should be easily and directly explainable to all participants on the customer side and on the vendor side, to remove potential for miscommunication. 3. Per square foot a. Raw square foot b. Rentable square foot c. Raised floor square foot d. Useable square foot 7. Adherence to market best practices The metric should allow for clear and repeatable processes, success metrics, and standards for purchasing that can be directly compared with other organizations, and with industry benchmarks. 4. Proprietary metrics: a. Per virtual data center used to denote a bundle of space, power, and bandwidth. b. Per cage / half cage used to denote a VDC of 4 or 2 standard racks respectively 3
5. Per kilowatt defined as sufficient space to house 1 kilowatt of power draw at peak utilization Checklist vs. Cost Analysis Goals can be established for previous deals or for other options on the market because no correlation is established between quantities and prices. 2. Allows spec change projections: no. Each change in spec must go back to vendor(s) for pricing. 3. Reflects vendor cost structure: yes. 4. Is vendor neutral: yes Strengths / Weaknesses of Each Approach Per Deployment Per-deployment is a price model favored by larger vendors for whom data center is a small piece in an overall managed services or full IT outsourced solution. Its biggest strengths are its brevity for an executive presentation and its ability to compare offers for the same are the possibility that the vendor did not properly size the environment as well as the need to re-run the RFP / RFQ for each small design change. Here's how it matches up against each criterion: 1. Apples-to-apples comparable: partial. Although offers between vendors in a deal can be compared, no benchmarks 5. Can translate into other models: no. No correlation is established between quantities and prices 6. Minimizes misunderstandings: yes (assuming spec is thoroughly defined) 7. Follows best practices: no. Knowledge from one exercise cannot be reused in others Per Rack Per rack (or cabinet) is the most common model for carriers and carrierneutral hubs, dating back from the time when all rack deployments used the same power draw. In fact, vendors in many non-us markets still enforce a standard rack size at a density that's incompatible with modern layouts. Its primary strength 4
is its universal appeal you will never have to explain to a vendor what you mean by "I need 10 racks." It weakness is that "10 racks" now has many more underlying meanings than before, creating an infinite array of ways not only to have genuine misunderstandings but to be intentionally misled by the other party. In practical terms, it tends to break down completely at densities of more than 3kw-4kw, creating artificial charge metrics such as "50 racks (20 useable)." 1. Apples-to-apples comparable: partial. Although offers between vendors in a deal can be compared, no benchmarks can be established for previous deals because racks can contain different amounts of power, which will affect price drastically. 2. Allows spec change projections: partial. Changes in quantities can only be projected provided the same power load per rack remains. 3. Reflects vendor cost structure: no. Most data center operational costs are driven by power / cooling, not space. 4. Is vendor neutral: yes 5. Can translate into other models: partial. Per kilowatt pricing can be established if the power spec is provided; however per-square foot pricing cannot be inferred easily. 6. Minimizes misunderstandings: no. Vendors are selectively blind to power requirements and price in a standard rack, hoping to get through initial down select. Only after moving to final negotiations do they announce, Well, we didn t realize you would need to use a non-standard rack with a non-standard price. Furthermore, some vendors offer taller racks, creating an issue where both quantities and prices change as equipment can be consolidated into fewer, more high-priced racks. 7. Follows best practices: no. Major organizations have a variety of rack configurations ranging from mid-range computers to high-density blades and cannot meaningfully compare cost/pricing with external vendors or market benchmarks. Per Square Foot Per square foot is the legacy real estate metric, so you will most often see it from vendors whose core competency is real estate rather than co-location or data center services. It is practically endemic 5
in in-house data center accounting (especially in the financial services vertical). Its primary strength is in the ability to work from layout diagrams if your design requires a specific physical configuration. Its biggest weakness is its inability to handle the true cost drivers of the data center environment today, which is power consumption rather than space usage in the vast majority of cases. Not all data center space is created the same, and therefore your predetermined layout may either be suboptimal or impossible with some of your providers. 1. Apples-to-apples comparable: no. Not only can benchmarks from past deals not be used due to differing power densities, even offers in the same deal cannot be evaluated meaningfully because both the quantities and unit rates of the services included change from sq ft offer to sq ft offer. 2. Allows spec change projections: partial. Changes in quantities can only be projected provided the same power load relative to space used remains. 3. Reflects vendor cost structure: no. Most data center operational costs are driven by power /cooling, not space. 4. Is vendor neutral: yes 5. Can translate into other models: partial. Per kilowatt pricing can be established if watts per square foot are provided. However, per-rack pricing cannot be inferred easily standard rules of thumb for converting square feet into racks become useless at densities above 3kw / rack. 6. Minimizes misunderstandings: no. The need to change both quantity and unit price simultaneously has historically proven highly confusing to vendors and does not allow for an easy definition of what is being requested (a 500 square foot cage if you re a high density vendor; a 750 square foot cage if you re a medium density one; and a 1,000 square foot cage for low density is obviously suboptimal). Furthermore, the very definition of a square foot can vary (especially in large purchases): a. Raw square footage entire building space including walls, elevators, etc. b. Rentable square footage removes building infrastructure such as elevator shafts c. Raised floor square footage removes office space d. Useable square footage removes support columns, major 6
walkways, space for UPS /PDU, etc. 7. Follows best practices: no. Major organizations use a variety of data centers with drastically different peak densities per square foot and cannot create a meaningful comparison metric across locations and against industry benchmarks. Proprietary Metrics Proprietary metrics have the same disadvantages as per-rack / per-square foot pricing, plus lack of vendor neutrality, making them a non-starter for most of our clients. Per Virtual Data Center (VDC) This metric originated with Exodus (now Savvis), and was adopted by other vendors for a period of time. However, it represents a product definition as a service description and, for Savvis, as a market differentiator. We generally only see it in legacy situations, where a contract has not been closely reviewed in several years. Per Cage Cage is a metric that can denote either a fixed or variable amount of data center space: Savvis used "cage" and "half cage" to denote a specific type of VDC configuration with sufficient space for 4 and 2 racks respectively. Some vendors also use "cage" as the unit in a per-deployment pricing model, where each distinct footprint appears on the client's bill as a cage. This ambiguity in the usage of the term gives buyers yet another reason to seek quotes denominated in a different set of units. Per Kilowatt This approach was pioneered in largescale buyers of entire wholesale data centers, and by companies who build their own facilities. Today it is propagating in the market largely through the efforts of wholesale providers such as Digital Realty Trust and CoreSite as well as RampRate's own work with our clients. This approach has many advantages as a simple metric that can measure the footprint at almost any type of facility with just one number. Its primary disadvantage is that, although each data center operator worth its salt knows its cost per kilowatt, many neglect to train their salespeople 7
on properly interpreting these requests and they do not adapt their price sheets or billing systems from legacy configurations. We expect per-kilowatt pricing to become dominant (or at least universally understood in the same way that rack pricing is) in the US in the next three years, with other world locations trailing a little behind. 1. Apples-to-apples comparable: yes. Each offer in a given transaction can be compared to other offers in the same transaction, as well as to other offers and other transactions (except ultra-low density deployments such as space for tape libraries or aged mainframes). 2. Allows spec change projections: yes. Per-kilowatt pricing most closely matches vendors internal engineering calculations governing cost and allows most accurate projections of cost increases /decreases based on changed specifications regardless of change type. 3. Reflects vendor cost structure: yes. Directly ties to key cost drivers of power and cooling 4. Is vendor neutral: yes 5. Can translate into other models: partial. Price per rack is typically identified through definition of site specs, but price per square foot is not considered essential to the quote process, and is occasionally omitted because it needs vendor-specific density information to calculate. 6. Minimizes misunderstandings: yes. Since the specific terminology of perkilowatt pricing is not used by the majority of vendors in other contexts / models, if the client defines the metric thoroughly, no opportunity for confusion arises. 7. Follows best practices: yes. RampRate has worked with major media and financial services companies who have already adopted per-kilowatt as a metric for tracking data center space costs, or are planning to do so. This group includes Intel and Digital Realty Trust. Key Questions on Data Center Cost Metrics Q: If neither we nor our vendors use per-kilowatt pricing, why would we use a model that most vendors don t use in their sales process / billing A: For the purposes of requesting a quote, the lack of familiarity can be very 8
beneficial. If forces you and the vendor to speak the same language, and removes the risk of confusion from what we ve always done that has built up over years of conflicting definitions and adherence to outdated schematics (e.g. that all racks use a standard 110V / 20A circuit, or 1.5kw of power). If the vendor wants your business and be assured that they do they ll figure out a way to meet your requirement in the sales and billing process. And we can almost guarantee that all vendors are using per-kilowatt pricing for some portion of their customers, even if they haven t started doing so with you. Q: Our experience has been that price per square foot does not vary with the vendor s data center density (maximum watts / square foot) meaning that the lower the vendor s density, the higher the total cost. However, per-kilowatt pricing implies a direct relationship total cost stays constant, but as density decreases, the cost per square foot should also decrease. A: Vendors participating in competitive markets recognize that buyers make decisions based on the bottom line. If pricing per square foot were, in fact, constant regardless of power density, then high density data centers would always be price leaders in terms of total cost. This is not the case. In wellnegotiated deals in our experience, when faced with a competitor reducing costs through increasing density, lowdensity vendors can and do adjust their pricing to match the bottom line quotes on the table. While there are some unrelated business drivers (strategic providers and MSPs like IBM and AT&T happen to run some very low density data centers and have high prices), we have found that high density does carry a premium and low density a discount. Furthermore, for the purposes of projecting a budget for a specific set of services, using a methodology where total pricing is contingent on vendordriven variables such as data center density does not appear appropriate or meaningful. Building a budget or benchmark that said, Your costs for this configuration will be $X, unless you choose a lower or higher density data center than our default assumption would not serve the purpose of truly reflecting the breadth of pricing available in the market. Q: Does per-kilowatt price modeling work for all deployments? 9
A: Per-kilowatt price modeling works best for medium to high densities. Low densities often do still have space as a core cost driver, which skewing your price. Ultra-high densities may require vendors to undertake new CapEx in cooling equipment (e.g. water or CO2 cooling) that can increase per-kilowatt pricing. Q: How would we request a quote for 1 kilowatt of space? A: There are two potential approaches. One is to directly say, We need sufficient space to house X kilowatts of power. The other is to ask for pricing for X racks drawing Y kw of power each. used. The space pricing methodology above uses peak rather than average (typically 80% of breaker and 133% of average) power capacity. Nevertheless, multiple organizations, both on the vendor side and on the client side, have not only tracked power based on peak kilowatt, but rolled space and power pricing into one total figure per kilowatt (or, more typically per megawatt) when buying at scale. Q: Can we discuss this with RampRate in more detail? A: To schedule an in-depth conversation, please contact us at the following: Q: Should we use per-kilowatt cost metrics for power as well? A: Although it is not RampRate s preferred approach, per kilowatt cost modeling (to remove confusion, potentially referred to as kilowatt-month as opposed to kilowatt-hour) is also possible for power. One nuance to keep in mind is that, when forecasting for usage-based power, you should use average kilowatts drawn (typically 60% of breaker capacity and 75% of peak capacity), which can then be multiplied by the number of hours in a month to understand kilowatt hours of power 235 Montgomery, Suite 1017 San Francisco, CA 94104 (415) 373-0673 www.ramprate.com or send an email to info@ramprate.com. 10