operation research homework
This article was downloaded by: [132.174.254.97] On: 25 February 2017, At: 11:22 Publisher: Institute for Operations Research and the Management Sciences (INFORMS) INFORMS is located in Maryland, USA Interfaces Publication details, including instructions for authors and subscription information:
http://pubsonline.informs.org Optimizing Capital Investment Decisions at Intel Corporation Karl G. Kempf, Feryal Erhun, Erik F. Hertzler, Timothy R. Rosenberg, Chen Peng, To cite this article:
Karl G. Kempf, Feryal Erhun, Erik F. Hertzler, Timothy R. Rosenberg, Chen Peng, (2013) Optimizing Capital Investment Decisions at Intel Corporation. Interfaces 43(1):62-78. http://dx.doi.org/10.1287/inte.1120.0659 Full terms and conditions of use: http://pubsonline.informs.org/page/terms-and-conditions This article may be used only for the purposes of research, teaching, and/or private study. Commercial use or systematic downloading (by robots or other automatic processes) is prohibited without explicit Publisher approval, unless otherwise noted. For more information, contact [email protected]. The Publisher does not warrant or guarantee the article’s accuracy, completeness, merchantability, fitness for a particular purpose, or non-infringement. Descriptions of, or references to, products or publications, or inclusion of an advertisement in this article, neither constitutes nor implies a guarantee, endorsement, or support of claims made of that product, publication, or service. Copyright © 2013, INFORMS Please scroll down for article—it is on subsequent pages INFORMS is the largest professional society in the world for professionals in the fields of operations research, management science, and analytics.
For more information on INFORMS, its publications, membership, or meetings visit http://www.informs.org Vol. 43, No. 1, January–February 2013, pp. 62–78 ISSN 0092-2102 (print) —ISSN 1526-551X (online) http://dx.doi.org/10.1287/inte.1120.0659 © 2013 INFORMS Optimizing Capital Investment Decisions at Intel Corporation Karl G. Kempf Decision Engineering Group, Intel Corporation, Chandler, Arizona 85226, [email protected] Feryal Erhun Management Science and Engineering, Stanford University, Stanford, California 94305, [email protected] Erik F. Hertzler Technology Manufacturing Engineering, Intel Corporation, Chandler, Arizona 85226, [email protected] Timothy R. Rosenberg Financial Enterprise Services, Intel Corporation, Chandler, Arizona 85226, [email protected] Chen Peng Management Science and Engineering, Stanford University, Stanford, California 94305, [email protected]; currently at Google Inc., Mountain View, California, [email protected] Intel Corporation spends over $5 billion annually on manufacturing equipment. With increasing lead times from equipment suppliers and increasing complexity in forecasting market demand, optimizing capital investment decisions is a signi cant managerial challenge. In response to this challenge, we developed a capital supply chain velocity program for ordering, shipping, and installing production equipment. At the core of this veloc- ity program is a new and additional procurement framework that enables Intel to purchase options from its equipment suppliers for a faster delivery of some equipment. The framework seamlessly combines statistical forecasting with Monte Carlo simulation and stochastic programming to determine the number of options Intel should procure and exercise, and it includes built-in scenario and sensitivity analysis capabilities to support Intel's contract selection, options reservation, and equipment procurement decisions. The velocity program and the framework provided Intel with hundreds of millions of dollars in cost savings and at least $2 billion in revenue upside during a recent period of global economic crisis.
Key words : capital-intensive industries; capacity expansion; dual sourcing; expedited equipment lead times; option contracts; forecast revision; stochastic programming; Monte Carlo simulation; isopro t analysis. I ntel Corporation was founded in July of 1968 by Robert Noyce, coinventor of the integrated circuit, and Gordon Moore, a pioneer in the development of manufacturing processes for producing integrated circuits in high volume. Intel has topped the list of the world's largest semiconductor companies since 1992, a position it maintained in 2011 with sales of $54 billion. Intel generated this revenue by competing in a wide range of markets, including servers, desktop and laptop computers, ultramobile devices (e.g., tablets and smartphones), and embed- ded devices in vehicles, homes, and medical equip- ment. This broad set of markets is consistent with Intel's corporate vision: This decade we will create and extend computing technology to connect and enrich the lives of every person on earth. The markets for these products have changed dra- matically over the years, with demand in some markets growing exponentially, in other markets dis- appearing completely, and in new markets arising 62 THE FRANZ EDELMAN AWARD Achievement in Operations Research Downloaded from informs.org by [132.174.254.97] on 25 February 2017, at 11:22 . For personal use only, all rights reserved.
Kempf et al.:
Optimizing Capital Investment Decisions at Intel Corporation Interfaces 43(1), pp. 62–78, © 2013 INFORMS 63unpredictably. Each new and innovative product enters an increasingly complex and dynamic set of markets. Intel has supported this market evolution with technical innovations in line with Moore's Law (Moore 1965), which predicts a repeating stepfunction improvement of manufacturing processes, such that the number of transistors per unit area of an inte- grated circuit doubles every 24 months. Although this ongoing improvement of manufac- turing processes has been the engine driving Intel's success in everchanging markets, it repeatedly forces very dif cult capital investment decisions because each stepfunction improvement requires the purchase of new and improved manufacturing equipment. As Intel's manufacturing process technology increases in complexity so does equipment complexity, which translates to longer lead times and higher prices from Intel's many suppliers. However, as Intel's markets increase in number and size, so does the complexity in forecasting the highly uncertain and rapidly evolv- ing demand. Furthermore, Intel is continuously chal- lenged by new competitors resulting in pressure to reduce prices and ultimately lower costs to compete effectively and ef ciently. Thus, balancing increas- ing equipment lead times against increasing forecast- ing complexity and increasing equipment cost against increasing product pricing pressures are at the core of Intel's decision problem. In the following sections, we present Intel's his- torical approaches to this decision problem. We then describe a new approach developed through collabo- ration between practitioners at Intel and researchers at Stanford University. In this new approach, the Figure 1: The DMEP project history is shown as major events (arrows) plotted over time against manufacturing process life cycles designated by thick horizontal bars with ovals at peak production, including the transistor feature size they enable expressed in nanometers (nm). majority of the equipment required is procured from each supplier with contracts specifying stan- dard delivery lead times; a second mode is enabled through options contracts to procure a few pieces of equipment with reduced lead times. We call this new approach dual-mode equipment procure- ment (DMEP). It seamlessly combines several oper- ations research and management science (OR/MS) techniques, such as statistical forecasting with Monte Carlo simulation and stochastic programming, to sup- port Intel's strategic, tactical, and executional procure- ment decisions. Before reporting the broad positive impact on Intel and its equipment suppliers that has resulted from implementing DMEP, we discuss solu- tions to the cultural issues we faced during the DMEP implementation. The timeline in Figure 1 shows the major events in the ongoing project.
Intel's Initial Approaches Historically, Intel attempted to balance capacity and market risks in an intuitive but inef cient manner.
Given the choice between risking a few hundred million dollars in cost by having too much equipment and risking a few billion dollars in revenue by being short on capacity, the decision to carry excess capac- ity was intuitively obvious. This was an appealing approach because early excess orders could be with- drawn, incurring a cancellation fee as demand fore- casts became clearer over time. Intel tolerated some idle equipment to assure customer goodwill, espe- cially because the highest revenue in a product's life cycle is realized when it is introduced to the market on a new manufacturing process.45 Year 65 45 32 130 22 90 Manufacturing process designation 2001 02 03 04 05 06 07 08 09 10 11 12 13 14 Phase 1 (lithography equipment) Phase 2 (all equipment) Initial use of the “options on lead time” idea Capital supply chain velocity program Stanford direct involvement Downloaded from informs.org by [132.174.254.97] on 25 February 2017, at 11:22 . For personal use only, all rights reserved.
Kempf et al.:
Optimizing Capital Investment Decisions at Intel Corporation 64 Interfaces 43(1), pp. 62–78, © 2013 INFORMSValue of a New Risk-Management System:
Data Guiding Intel's Path to the Future In early 2000, operations researchers in Intel's decision engineering (DE) group, which is responsible for deci- sion support tools and methods, in close collaboration with industrial engineers and nancial analysts in Intel's technology manufacturing engineering (TME) group, which is responsible for sourcing and procur- ing equipment, began to question this approach to balancing risk. Analysis of historical data on the pre- vious ve manufacturing processes showed charges for cancelling equipment orders averaged $50 million per manufacturing process transition. Furthermore, the total cost of purchasing excess equipment ranged from $200 million to $400 million per transition.
Although Intel was never short of capacity, the com- pany estimated it had spent at least $1.25 billion (5 $50 million C5 $200 million) in managing the per- ceived risk. The DE and TME team recommended the development of a new risk-management system to address the double challenge of (1) including a more data-driven risk-management methodology, and (2) reducing the overall cost of managing the risk. Even with this hefty price tag, equipment suppli- ers were shouldering too much of the forecasting risk. This was strategically unacceptable, because Intel relies heavily on its suppliers for equipment inno- vations, impacting planned manufacturing processes ve to 10 years into the future. Any new risk-manage- ment methodology had to be a win-win situation for all involved. Ordering excess equipment with subsequent can- cellation was also unacceptable from a tactical perspective. Given the competitive nature of the semiconductor industry, if a supplier built a piece of equipment for an Intel order that was subsequently cancelled, the equipment was available immediately to Intel's competitors. Intel had paid a cancellation fee for an overforecasting error, enabling a competi- tor to leapfrog the 12- to 18-month lead-time queue to remedy an underforecasting error. Any new risk- management methodology had to avoid providing advantages to Intel's competitors.
The analysis provided an insight for developing a new approach. Figure 2(a) shows a detailed pro le of a typical manufacturing process life cycle as a detailed expansion of one of the manufacturing process bars in Figure 2: The upper diagram (a) shows the stepwise installation of capac- ity as demand ramps up, with the earliest and latest points in time to order equipment with a four-quarter lead time. The middle diagram (b) shows a high-demand forecast error four quarters and longer before demand peaks, resulting in a large gap between the upper and lower forecast con- dence bounds. The lower diagram (c) shows the dramatic reduction in forecast error within four quarters of the demand peak, including a much smaller gap between the upper and lower forecast con dence bounds.Mean –12 –10 – 8 – 4 0 2 4 6 8 10 % Demand forecast error Lead time Lead time –6 –2 Time relative to peak demand (quarters) Demand forecast volume (units) Upper boundUpper bound Mean –12 –10 – 6 – 4 – 2 0 2 4 6 8 10 Lead time Lead time –8 Time relative to peak demand (quarters) % Demand forecast error Demand forecast volume (units) Lower boundLower bound (c) (b) (a) Demand forecast volume (units) Time relative to peak demand (quarters) 0246810 –2 –4 –6 –8 –10 –12 Capacity installation stepDemand forecast mean Last equipment Lead time Lead time First equipment Downloaded from informs.org by [132.174.254.97] on 25 February 2017, at 11:22 . For personal use only, all rights reserved.
Kempf et al.:
Optimizing Capital Investment Decisions at Intel Corporation Interfaces 43(1), pp. 62–78, © 2013 INFORMS 65Figure 1. A new factory that costs in excess of $6 billion to build and equip may contain as many as 2,000 indi- vidual pieces of equipment of approximately 75 types and purchased from over three dozen major suppli- ers. Because of the inherent complexity of installing and testing this much equipment, production capac- ity is added incrementally in steps of 5 to 15 percent of the target total capacity. Initially, the minimal set of equipment to execute the entire manufacturing pro- cess is put in place, and the factory begins to pro- duce a saleable product. The next set of equipment is then added to realize another stepfunction increase in capacity. Increment by increment, the factory achieves full capacity, keeping pace with a growing market for its products. Because capacity is installed in incre- ments over eight quarters, assuming a four-quarter lead time, equipment orders are placed between 12 and four quarters before the demand peak. Figure 2(b) shows typical data for demand fore- cast volume in units of product and absolute demand forecast error; volume is computed as a percentage of the demand forecast by comparing the forecast at peak demand with the actual demand realized. This gure displays a typical demand forecast, including the mean volume line and con dence intervals, show- ing all equipment being ordered in periods of very high demand forecast error. Figure 2(c) shows the key insight taken from the extended percentage of demand forecast error line. Two quarters prior to peak demand, the forecast error is reduced by at least half. Based on this analysis, Intel clearly had two alterna- tives moving forward. A more accurate forecast fur- ther in advance of the demand peak would make capital investment decisions easier but is practically impossible to realize. Alternatively, shortening the lead time of some equipment to two quarters or less is imaginable. We enabled the latter approach through option contracts to procure a few lithography scan- ners with reduced lead time as a proof of concept.
A Pilot Study: Option Contracts for Lithography Equipment Scanners, which cost tens of millions of dollars per tool, are among Intel's most expensive equipment types. Initial implementation was a rudimentary con- tract heuristically addressing option cost and quan- tity. In the speci c case of scanners, roughly half the lead time is used in growing the lens, which is at the heart of the machine. The cost of an option was sim- ply the cost of growing a lens in advance. The bene t of the option was a 50 percent reduction in lead time once a rm order was placed. The number of options required was established by dividing the difference between the upper and lower bounds of the demand forecast in Figure 2(b) by the capacity of an individual scanner. This approach promised more effective risk management based on data instead of intuition, with a reduced chance of aiding the competition. Presentations to Intel management showing the estimated $1.25 billion spent on managing perceived risk for the previous manufacturing processes and explaining the concept of more effective risk mit- igation won management approval. Management authorized the TME team to test the approach on scanners for the next manufacturing process transi- tion to 130 nm (see Figure 1). Management at the scanner supplier was intrigued by the concept and agreed to participate after setting an upper limit on the number of options that it could purchase based on the supplier 's internal capacity constraints. After Intel and the supplier agreed to that number, a monthly decision meeting between TME and key managers from Intel manufacturing was convened to decide to exercise or defer the option. Although the nancial impact was limited to a few tens of millions of dol- lars, the experience gained in collaborating with the supplier and participating in the monthly option exer- cise meetings during 2002 and 2003 was invaluable (Vaidyanathan et al. 2005). Even personnel who were initially skeptical saw the potential. The use of scan- ner options was extended through 2008 to manage the subsequent two process transitions (130 nm to 90 nm to 65 nm, as Figure 1 shows).
Intel at Crossroads: Scaling the Pilot Project to All Equipment Types In early 2007, Intel senior management conducted a benchmarking study to compare Intel's capital supply chain performance against competitors and other industries. One nding called Intel's over- all capital investment performance into question.
This motivated the initiation of the capital supply chain velocity program for ordering, shipping, and installing production equipment, a program that tar- gets increasing the agility of the capital equipmentDownloaded from informs.org by [132.174.254.97] on 25 February 2017, at 11:22 . For personal use only, all rights reserved.
Kempf et al.:
Optimizing Capital Investment Decisions at Intel Corporation 66 Interfaces 43(1), pp. 62–78, © 2013 INFORMSsupply chain. The velocity program included extend- ing the implementation of our methodology from managing scanner risk to managing risk for all the equipment types used in the semiconductor manufac- turing process. To increase the expertise on the project and shorten the time to solution, Intel engaged our collaborators of long standing at Stanford University.
This collaboration led to the development of the new DMEP framework, which provides Intel with an improved methodology with sound theoretical underpinnings for determining the number of options needed, the valuation of the options, and the appro- priate timing for exercising the options (or not) as forecasts evolve.
A Systematic Approach to Intel's Problem:
DMEP Framework Intel's capacity planning and procurement process materializes in three stages. In the rst stage, several Figure 3: The graphic shows the three stages of the DMEP procedure. Intel and its supplier rst agree on the parameters of the base and exible (BASE and FLEX) supply modes (contract negotiation stage). Intel then reserves the total equipment procurement quantities BT and FT from these modes (reservation stage). Finally, Intel orders equipment from the two supply modes, given the latest demand forecast and the realized demand information from the market (execution stage). years in advance of equipment procurement, Intel writes umbrella contracts with its suppliers. These umbrella contracts outline the relationship between the two parties but do not specify all operational details. In the second stage, two to three years in advance of equipment procurement, Intel and its suppliers write amendments to these umbrella con- tracts. The amendments detail operational factors, such as forecasted procurement quantities. In the third stage, as the new manufacturing process ramps up in factories, Intel exercises these contracts and pro- cures equipment incrementally over time, as we dis- cussed above. Intel's previous methods of dealing with this pro- cess considered these stages independently and relied on intuition and simple heuristics for decisions.
However, these three stages are related and should be considered together. The amount of exible capac- ity Intel reserves depends on the possible demandContract negotiation problem (3–5 years prior to peak) Determine the best (lead time, price) combination from the contract menu Base reservation quantityB T Reservation problem (3 years prior to peak) Determine the total equipment amount to reserve through both supply modesFlexible reservation quantityF T Base mode (withL b) Base orderB i Execution problem (every decision period)Flexible mode (withL f) Flexible orderF i Determine how much equipment to order from both reservation poolsForecast revision Downloaded from informs.org by [132.174.254.97] on 25 February 2017, at 11:22 . For personal use only, all rights reserved.
Kempf et al.:
Optimizing Capital Investment Decisions at Intel Corporation Interfaces 43(1), pp. 62–78, © 2013 INFORMS 67realization scenarios, and setting the option quan- tity without considering these scenarios leads to inef ciencies.
Thus, we developed a three-stage decision support tool (see Figure 3) that enables Intel to procure equip- ment from its supplier using two supply modes with complementary lead times and prices: a base mode and a exible mode (BASE and FLEX, respectively, in the remainder of this paper). BASE, the regular pro- curement mode (see the left side of Figure 3), is less expensive but has a longer procurement lead time L b; FLEX (see the right side of Figure 3) is more expensive but has a shorter procurement lead time L f.
During the contract negotiation stage, several years before the adoption of a new manufacturing pro- cess and the procurement of the necessary equipment, Intel and its supplier agree on the parameters of the supply modes. They negotiate lead times and prices (a reservation price and an execution price) for the BASE and FLEX modes. Access to the FLEX mode allows Intel to react to the market conditions later with better demand information; however, Intel pays more for shorter procurement lead times; choosing the right lead time and cost combination in this stage is crucial. Note that, given the complex nature of the equipment and its precision required by Intel, sourc- ing the same equipment from two different suppli- ers is highly unlikely. Thus, Intel uses a sole-sourcing strategy for each process and equipment pair. Our approach is based on sourcing from two different modes of the same supplier instead of sourcing from two different suppliers. During the reservation stage, several quarters before the planning horizon starts, Intel procures options (i.e., reserves the total equipment procure- ment quantities) BT and FT from the BASE and FLEX supply modes at the previously agreed-upon unit reservation prices. The reservation payment enables the supplier to obtain capital to prepare its produc- tion capacity. This helps distribute the inherent risk of capacity expansion between Intel and its suppliers, and Intel's capacity commitments guarantee a speci c pro t for the suppliers. During the execution stage in each period n, Intel orders equipment from the two supply modes at respective unit execution prices, given the latest de- mand forecast and the realized demand information from the market. Early equipment procurement reduces the potential risk of having unsatis ed demand; however, it also leads to higher equipment holding costs and high opportunity costs if the equip- ment remains idle.
DMEP is a multistage decision hierarchy that guides Intel during different phases of equipment procurement. It provides decision support in ap- proaching questions such as: How can Intel quickly evaluate different FLEX options during contract nego- tiations? Under what circumstances does the FLEX mode create value for Intel? How much of the total capacity should be reserved through the FLEX mode?
When should this capacity be exercised? With its three stages, DMEP enables Intel man- agement to plan and operate an agile capacity sup- ply chain that adapts effectively to changing demand conditions. To provide such guidance, the decision support system seamlessly combines statistical fore- casting with Monte Carlo simulation and stochastic optimization and includes built-in scenario and sensi- tivity analysis capabilities.
Literature Review The DMEP procedure implements a nite-horizon, periodic-review, dual-source capacity expansion model by building on and extending three streams of literature: dual-source inventory management, forecasting, and strategic capacity decisions. It is well known in the dual-source inventory lit- erature that as long as the lead times of the two sources are consecutive, a modi ed base-stock pol- icy is optimal for these sources (Daniel 1963, Fukuda 1964). However, when the lead times are noncon- secutive or more than two sources exist, this policy is no longer optimal (Feng et al. 2006, Yazlali and Erhun 2009). Despite the existence of literature on the dual-source inventory problem, research concern- ing the dual-source capacity procurement problem is scarce. The exception is Chao et al. (2009), in which the authors investigate the problem with lost sales.
They establish that a base-stock policy is optimal for the exible source; furthermore, when the capacity obsolescence rate is deterministic, a base-stock policy is also optimal for the base source.
DMEP's contract negotiation stage builds on the literature on supply contracts, which Cachon (2003)Downloaded from informs.org by [132.174.254.97] on 25 February 2017, at 11:22 . For personal use only, all rights reserved.
Kempf et al.:
Optimizing Capital Investment Decisions at Intel Corporation 68 Interfaces 43(1), pp. 62–78, © 2013 INFORMSreviews. The negotiation of price and lead-time com- binations is also related to the literature on pric- ing and lead-time quotations. This literature often assumes that the buyer 's order quantity is a function of the price and lead time quoted by the supplier; thus, it focuses on the supplier 's optimal price and lead-time decision (Liu et al. 2007).
Our contributions to these streams of literature are as follows. For the execution module, we study dual- source capacity procurement with backlogging (i.e., late shipments to Intel customers, which we penal- ize with decreasing pro t margins and a high penalty cost at the end of the planning horizon) and show that the optimal ordering policy lacks structure even in the simplest setting (Peng et al. 2012). Thus, we con- tribute to the literature by showing that a dynamic dual-source capacity expansion problem with back- orders is categorically different and more complex than its inventory counterpart and the dynamic dual- source capacity expansion problem with lost sales.
Our work also contributes to the literature on the capacity expansion problem (reservation stage), which has been extensively studied under single sourcing (Van Mieghem (2003) provides a review of this litera- ture) by considering a dual-source multiperiod capac- ity expansion problem. Finally, our contract negotia- tion module contributes to the literature by taking the buyer 's perspective; we provide the buyer with a deci- sion support tool that can be used while negotiating prices and lead times with suppliers. Our theoretical results for the dual-source capac- ity procurement problem with backlogging show that a dynamic programming-based method would be highly inef cient to deal with Intel's problem with nonconsecutive lead times and comprehensive forecast-updating mechanisms. Therefore, we rely on an alternative approach that combines stochastic pro- gramming with Monte Carlo simulation, as we dis- cuss in the next section.
Modules of the Dual-Mode Equipment Procurement Framework To map the framework presented in Figure 3, the DMEP heuristic consists of the same three layers:
the outermost is the contract negotiation layer, which identi es indifference curves of lead-time and price combinations so that Intel can pick the best alternative from the contract menu; the middle is the reservation layer, which calculates the optimal equipment reser- vation quantities BT and FT for the two supply modes; and the innermost is the execution layer, which deter- mines the equipment order quantities from the two supply modes in each period. Before elaborating on each of these three layers in more detail, we rst con- sider the demand forecast revision process, which is at the core of optimizing capital investment decisions.
Forecasting Module Figure 4 graphically describes the demand forecast- updating process at Intel. The left panel shows how the mean forecast and forecast error for a certain period (e.g., period 0) evolve as we approach period 0. The solid curve represents the mean forecast evolu- tion path, and the interval between the dashed curves denotes the forecast error range, which shrinks as the forecasting lead time decreases. That is, the forecast accuracy for a certain period's demand improves as we move closer to the period. This improvement is solely a function of time; moreover, we see a dramatic decrease in forecast errors two quarters prior to peak demand (see Figure 2 and the discussion that follows).
The right panel of Figure 4 depicts how we use these observations to forecast all future periods at a given period (e.g., period 0): the solid curve represents the mean forecast and the dashed interval denotes the forecast error range, which diverge as the forecasting lead time increases. That is, the forecasting accuracy decreases as we forecast further into the future. We consider several factors in Intel's forecasting process. Each business unit is responsible for its Figure 4: The graphic illustrates Intel's demand forecast updating pro- cess. Forecast error for an anchor period's demand decreases as we move closer to the period, as the left panel shows. We can use this observation to forecast all future periods' demand at an anchor period, as the right panel shows.00 –2 –4 Forecast errorRange forecastRange forecast Mean forecast Anchor period Anchor period Time Time 4 2 Downloaded from informs.org by [132.174.254.97] on 25 February 2017, at 11:22 . For personal use only, all rights reserved.
Kempf et al.:
Optimizing Capital Investment Decisions at Intel Corporation Interfaces 43(1), pp. 62–78, © 2013 INFORMS 69own demand forecast using historical sales data, new product introductions, geographic demand data, and seasonal factors. Intel also analyzes many economic data points (e.g., Semiconductor Industry Association (SIA) data for the whole market), holds discus- sions with customers at many levels, and studies the announcements and activities of competitors as part of its forecasting process. Because Intel is at least one generation ahead of its competition in manufacturing process technology, an added dif culty is estimating the market response to a new feature. Furthermore, software innovators and a host of others in the ecosys- tem create additional forecasting complexities. To capture Intel's forecasting process in a system- atic way, we modify the classical Martingale model of forecast evolution (MMFE) (Hausman 1969, Graves et al. 1986) by decomposing it into a mean evolution process and a forecast error evolution process. This decomposition is driven by the empirical observation of independent forecast mean and forecast error evo- lutions based on historical data from Intel. We assume that the mean forecast evolution of demand for a certain period nfollows a Markov process. Follow- ing Intel's forecast error evolution, we further assume that the forecast error is predictable and is uniquely determined by the forecasting lead time. This lat- ter modeling assumption helps us capture Intel's forecast-updating process in an easy-to-implement format. Especially when compared to the original multiplicative MMFE process, our approach avoids calculating cumulative distributions that involve the multiplication of several random variables. As such, our approach achieves a trade-off between analytical rigor and managerial applicability. Peng et al. (2012) provide the details of the forecasting process.
Execution Module The execution module is the core of the DMEP heuristic. Given the reservation quantities BT and FT , the execution module characterizes how Intel should place the BASE and the FLEX orders in each period.
At the beginning of each period, Intel obtains the lat- est demand realization and forecast updates for all future periods. Based on this information, all previ- ously placed BASE and FLEX orders and the back- log quantity from the preceding period, the execution module solves a stochastic program at each period to calculate the myopic optimal BASE and FLEX order quantities for the remaining periods. The objective function of the stochastic program (see the appendix) maximizes the expected pro t by considering the pro t margin, equipment pro- curement costs, and equipment holding costs. The expectation is taken over all future demand, and the execution cost is calculated when the orders arrive. Because our goal is to represent the strategic- and tactical-level capacity decisions, we suppress (1) the operational-level product inventory problem and assume that production in a period will not exceed demand and (2) supply variability. Because the strategic- and tactical-level capacity decisions are made years in advance, excessive uncertainty exists regarding the demand and supply (e.g., yields) pro- cesses. Therefore, we simpli ed the execution-level problem and concentrated on the most critical deci- sions to avoid including additional noise. Constraints guarantee that (1) sales in a given period cannot be larger than the demand or the supply; (2) Intel satis- es at least –percent of the incoming demand during each period after ful lling the backlogs; and (3) the BASE and FLEX orders are within their reservation limits and nonnegative. The execution module is a rolling-horizon algorithm; that is, in period m, only the orders that must be executed in period mwill actually be placed, and the rest of the decisions will be postponed to future periods when better demand information and forecasts will become available. We solve this stochastic program using the stan- dard sample average approximation method based on Monte Carlo simulation (Shapiro 2008). Assuming that demand is discrete and takes nitely many val- ues, the stochastic program can be converted into an ef cient linear program. Therefore, we are able to con- sider a large number of samples, which improves the accuracy of the solution. Hence, we do not have to rely on re nements of Monte Carlo simulation, such as importance sampling, that aim to reduce the vari- ance of estimation and improve its power.
Figure 5 illustrates the decision-making process.
Assume there are ND6 periods in the selling sea- son, the BASE mode lead time is L b D 4 periods, and the FLEX mode lead time is L f D 2 periods. At the beginning of the planning horizon, period ƒ3, the latest demand forecast information for period 1 toDownloaded from informs.org by [132.174.254.97] on 25 February 2017, at 11:22 . For personal use only, all rights reserved.
Kempf et al.:
Optimizing Capital Investment Decisions at Intel Corporation 70 Interfaces 43(1), pp. 62–78, © 2013 INFORMSFigure 5: The graphic illustrates a tabular demonstration of the execu- tion module, assuming L b D 4and L f D 2. At each decision epoch, new demand and forecast information become available to Intel: the grey cells provide the mean and error for the updated forecast made during season j for selling season i 4j < i 5, where Œj i denotes the mean of the forecast and ‘ iC— j— denotes the forecast error corresponding to a forecasting lead time of i C — j— periods; Dmddenotes realized demand; current period decisions to be executed; _ previously executed decisions; and grey current period decisions not to be executed.
period 6 is revealed. The algorithm runs the aforemen- tioned stochastic program to maximize the expected total pro t across the entire planning horizon under the current demand information, taking the six BASE orders ( B 1 to B 65 and six FLEX orders ( F 1 to F 65 as the decision variables. Once the optimal order quan- tities are obtained, Intel only needs to commit to B 1 at period ƒ3; it postpones the rest of the decisions ( B 2 to B 6 and F 1 to F 65 until later. At the beginning of period ƒ2, the updated forecast information is obtained; the rm then solves a new stochastic pro- gram to maximize the horizon-wide expected pro t under the newly obtained demand information, x- ing B 1 and taking B 2 to B 6 and F 1 to F 6 as the decision variables. Similarly, at period ƒ2 only the decision B 2 needs to be executed, and the rest of the decisions are left for later periods. Following this logic, B 3 and F 1 will be executed in period ƒ1, etc. This rolling-horizon decision-making process continues until period 4, when the nal order F 6 is committed. Note that start- ing from period 1, the actual realized demand is treated as part of the updated demand information and should be considered when calculating the total expected pro t. The execution module provides a solution that indi- cates when and how much to order from the two supply modes. It captures the evolution of demand information and enables timely decision making.
Unlike the rule-of-thumb approaches Intel previously relied upon, the necessity of the FLEX mode is eval- uated in each period of the planning horizon. We note that the execution module is myopic in the sense that it nds an ordering scheme based on the current information without considering the opportunity to make decisions contingent upon the actual realized demand at each stage. Fortunately, this disadvantage of myopia is mitigated by the rolling-horizon nature of the module. In practice, we have found the heuris- tic signi cantly outperforms unaided human judg- ment because of its forward-looking nature and the various trade-offs that it effectively captures.
Reservation Module As we explained above, letting the supplier bear most of the procurement risk is not an acceptable situation.
Therefore, DMEP's reservation module functions as a mechanism for risk sharing between Intel and its supplier. By paying an up-front reservation fee, Intel shares the risk of capacity building and installation with the supplier and enjoys the guaranteed delivery of equipment in return. Determining how much capacity to reserve from the two supply modes, especially the FLEX mode, is based on a trade-off between the reservation cost and the potential bene ts from the guaranteed exibility.
Speci cally, the optimal reservation quantities BT and F T are determined using a Monte Carlo simulation that considers various scenarios of the future demand pro les. Intuitively, if the future demand scenario involves no uncertainty, then exibility has no value because ordering from the expensive FLEX mode is never optimal. In contrast, if the future demand sce- nario is highly uncertain and the demand mean fore- cast is likely to be modi ed, then exibility has a high value and we should expect a higher level of reser- vation for the FLEX mode. In general, the reserva- tion quantities maximize the expected horizon-wide pro t over all possible demand scenarios. For each path, the reservation module calls the execution mod- ule to calculate the order quantities and the expectedPlanning horizon Selling season 2345 6 ( 1–3, 4)( 2 –3, 5)( 3 –3, 6)( 4 –3, 7)( 5 –3, 8)( 6 –3, 9) ( 1–2, 3)( 2 –2, 4)( 3 –2, 5)( 4 –2, 6)( 5 –2, 7)( 6 –2, 8) * * B4 B5 B6 B4 B4 B3 B2 B3 B5 B5 B6 B6 B3 B4 B5 * ** F3 F3 F3 F3 F4 F4 F4 F5 F5 F5 F6 F6 F6 F2 F2 F1 F2 F1 F4 F5 (Dmd1) B 1 B1 B1 B1 B1 B2 B2 B3 B2 B2 B3 B4 B5* * B6 F1 F1 F1 F2 F2 F3* F4 F5 F6 B6F6 1 ( 2 0, 2)( 3 0, 3)( 40, 4)( 5 0, 5)( 6 0, 6) ( 10, 1) ( 2 1, 1)( 3 1, 2)( 4 1, 3)( 5 1, 4)( 6 1, 5) ( 2 –1, 3)( 3 –1, 4)( 4 –1, 5)( 5 –1, 6)( 6 –1, 7) ( 1–1, 2) Downloaded from informs.org by [132.174.254.97] on 25 February 2017, at 11:22 . For personal use only, all rights reserved.
Kempf et al.:
Optimizing Capital Investment Decisions at Intel Corporation Interfaces 43(1), pp. 62–78, © 2013 INFORMS 71Figure 6: The graphic shows the DMEP modules and corresponding solution approaches. The decision support tool seamlessly combines statistical forecasting with Monte Carlo simulation and stochastic optimization to provide a solution for all three stages of Intel's problem, as Figure 3 displays.
horizon-wide pro t. We then choose the reservation quantities BT and FT that maximize the average total pro t across the entire planning horizon. The reservation algorithm helps Intel determine the optimal amount of equipment to reserve for both sup- ply modes. It builds on the philosophy of scenario analysis and chooses the reservation quantities that guarantee the maximum expected return to the rm.
By adjusting the mean evolution process (see the Fore- casting Module section), we can easily create different demand scenarios; hence, the reservation algorithm is applicable to a wide range of business settings.
Contract Negotiation Module During the contract negotiation stage, Intel and the equipment supplier determine the lead times ( L b, L f5 of the two delivery modes and the unit reservation 4r b1 r f5 and execution ( c b, c f 5 prices associated with these lead times. Our algorithm here involves a sen- sitivity analysis: for different (e.g., lead time, price) combinations, we run the reservation and execu- tion modules and obtain the corresponding expected horizon-wide pro ts. The decision maker can then choose the pair (e.g., lead time, price) from the con- tract menu that leads to the highest expected return.
The isopro t curves also help Intel identify inferior contracts from the menu of contracts offered by the supplier. The strength of this part of the heuristic is that it helps Intel make strategic-level decisions by considering potential tactical- and execution-level contingencies.
DMEP in Summary DMEP constitutes a stable decision support pyra- mid, where the managerial decisions are made in a top-down sequence (i.e., starting with the con- tract negotiation stage) whereas the decision support tool aids those decisions by evaluating the modules in a bottom-up order (i.e., starting with the execu- tion module). The decision support tool provides a solution that effectively includes all complexities of Intel's problem, such as contract negotiations, capac- ity reservations with options, dual-mode procurement with different lead times, and forecast updates. It is highly ef cient and it helps Intel executives evalu- ate their higher-level strategies by considering their impact on lower-level decisions. Furthermore, DMEP can be modi ed easily to include additional factors that may be relevant for other rms. For example, the execution-level algorithm can easily be modi ed to include (1) the choice of holding product inventory and (2) order bounds on either the total capacity that can be reserved from a mode or the capacity that can be exercised at a given period from a mode. We for- mulate the procurement problem as a linear program.
As such, DMEP is better suited to providing the frac- tion of orders from each mode rather than the actual procurement quantities. However, if the intention is to use DMEP to obtain the speci c order quantities, then the execution module can be generalized to an integer program. Figure 6 displays the modules of DMEP and the corresponding solution approaches and timeline for each module.
Implementation Conceptualization and Development of DMEP Intel content experts and the Stanford research team collaborated to complete the conceptualization and development of the framework and modules. The Stanford team completed primary development ofContract negotiation module:
Isoprofit analysis Reservation module:
Monte Carlo simulation Execution module:
Stochastic programming Forecasting module:
Martingale model of forecast evolution Downloaded from informs.org by [132.174.254.97] on 25 February 2017, at 11:22 . For personal use only, all rights reserved.
Kempf et al.:
Optimizing Capital Investment Decisions at Intel Corporation 72 Interfaces 43(1), pp. 62–78, © 2013 INFORMSthe framework and modules with frequent consulta- tion on the parameterization and necessary assump- tions. This team spent a considerable amount of time tuning DMEP. Together, we performed in-depth sen- sitivity analysis on several parameters that de ne Intel's business environment. Peng et al. (2012) pro- vide details of the scenario analysis and the parameter values. In particular, we considered various forecast errors, supplier pricing structures, lead-time com- binations, and service-level requirements for Intel products. We compared our results with historical actuals to validate the parameters used. After consid- ering thousands of scenarios for both the execution- and tactical-level problems, and discussing the results with Intel executives and planners, we came up with a range of 8 percent to 14 percent for the FLEX mode for most representative parameter settings at Intel. This range was our suggestion to Intel executives, with a target of 12 percent of total capacity sourced with the FLEX mode.
Transfer and Adoption of DMEP at Intel We completed the DMEP transfer and adoption in two phases. The rst phase consisted of knowledge transfer from Stanford to Intel and was performed in parallel with the software training of TME per- sonnel at Intel by Stanford researchers. The second phase involved internalization of the model by the TME group. This phase involved transformation of the Stanford implementation in a MatLab and CVX software package (run on an Intel Core 2 Duo proces- sor at 2.26 GHz, with a runtime of 2–3 hours) to the Intel implementation in an ILOG and Excel software platform using Excel's built-in randomization features (run on an Intel Core 2 Quad processor at 2.4 GHz, with a runtime of 15–20 minutes runtime). Intel personnel have extended the modeling methodology from the gross technology-sizing model used to establish BASE and FLEX percentages orig- inally requested of the Stanford team to a more detailed model at the individual supplier and spe- ci c equipment-type level. This has allowed TME to negotiate with each supplier based on the speci c cost and lead-time trade-off of each individual equip- ment type. DMEP drove the reservation quantities for the 45 nm to 32 nm manufacturing process transi- tion capacity procurement (see Figure 1) and is in use for Intel's current 32 nm to 22 nm process transition capacity procurement.
Implementation Challenges Addressed by the Project Beginning in the spring of 2007, Intel rolled out its vision of achieving the capital supply chain velocity program with DMEP to all of its supply chain trad- ing partners at its annual supplier day. Speci c sup- plier site visits by Intel senior management followed.
Initially, almost all suppliers stated that they could not achieve Intel's velocity goals without a signi cant increase in the cost of the equipment. After reaching consensus that demand variability impacts both sup- pliers and Intel, it was clear that forecast errors affect suppliers most. Intel asked each supplier to reevaluate hidden factory costs associated with managing this high variability. Through this process, opportunities for suppliers to shorten lead times were uncovered.
These opportunities would enable Intel to provide suppliers with more accurate demand forecasts. With DMEP, instead of asking equipment suppli- ers to make all their equipment in short durations, Intel was initially only asking for a small percentage to be made available with short lead times. To fur- ther incentivize its equipment suppliers, Intel agreed to collaboratively partner on making short lead-time equipment less disruptive. Examples include elimina- tion of duplicate inspections predelivery and postde- livery, and of assembling equipment at the supplier 's factory to disassemble, ship, and reassemble at Intel.
Now equipment is assembled only once at Intel.
Financial Bene ts to Intel Successfully addressing the technical and cultural issues resulted in applying DMEP to the transition from our 45 nm process to our 32 nm process targeted for initial production in early 2009 (see Figure 1). Intel implemented DMEP with 13 sup- pliers who represent the 48 most critical equipment types for the 32 nm manufacturing process transition.
Figure 7 shows the planned placement of BASE orders matched to the demand forecast issued in Q1 2008 and re ects the stepwise installation pattern shown in Figure 2(a). The initial BASE orders with a four-quarter lead time were placed in the rst three quarters of 2008Downloaded from informs.org by [132.174.254.97] on 25 February 2017, at 11:22 . For personal use only, all rights reserved.
Kempf et al.:
Optimizing Capital Investment Decisions at Intel Corporation Interfaces 43(1), pp. 62–78, © 2013 INFORMS 73Figure 7: The initial plan for the 45 nm manufacturing process to 32 nm manufacturing process transition was eight incremental BASE equipment orders to meet the Q1 2008 demand forecast. The dramatic market downturn in Q4 2008 and Q1 2009 led to reduced demand forecasts and the cancellation of BASE equipment orders 4 and 5.
to procure equipment for the initial 32 nm ramp in the rst three quarters of 2009. Unfortunately, over the next few months, the market declined sharply.
Figure 8 shows conditions in this period repre- sented by total booked orders (SIA Historical Billings Reports 2012). We include data from the previous Figure 8: Semiconductor Industry Association order data show the time- phased severity of the last four market downturns. three major market downturns for comparison. It is important to note that the 2008–2009 downturn was the steepest in recent memory and was within 10 per- cent of being the deepest downturn in memory. In Q4 2008 as the downturn started and a new demand forecast was issued, capacity procurement decisions had to be made about BASE order 4 (see Figure 7). Prior to implementing DMEP, Intel would have had two possibilities, based on the intuition and in uenced by history, that market recovery can take as long as two to three years (see Figure 8).
Many equipment orders could have been cancelled, incurring penalties two to three times above the $50 million average under normal circumstances.
However, the equipment orders could have been hon- ored and the equipment paid for and received. In this case, the equipment could have been warehoused or installed while Intel waited for a market upturn.
Either choice would have incurred depreciation of roughly 2 percent of the purchase price per month (approximately $12 million per month), the instal- lation choice incurring an additional cost to keep the equipment in idle or warm mode (approximatelyBASE: order 8 Q1/08 forecast BASE: order 7 BASE: order 5 BASE: order 6 BASE: order 1 BASE: order 2 BASE: order 3 BASE: order 4 Q4/08 reforecast Q1/09 reforecast Q1–08 Q2–08 Q3–08 Q4–08 Q1–09 Q2–09 Q3–09 Q4–09 Q1–10 Q2–10 Q3–10 Q4–10 Q1–11 Q2–11 1984–1986 1995–1996 2000–2002 2008–2009 10.00.0 – 20.0 –10.0 – 40.0 – 30.0 – 50.0 – 70.0 – 60.0 Volume percentage of downturn Quarters into downturn 7 6 5 4 3 2 189 Downloaded from informs.org by [132.174.254.97] on 25 February 2017, at 11:22 . For personal use only, all rights reserved.
Kempf et al.:
Optimizing Capital Investment Decisions at Intel Corporation 74 Interfaces 43(1), pp. 62–78, © 2013 INFORMS$10 million per month). The risk associated with buying the equipment was that the market would not come back and $500 million worth of equipment would be idle throughout its life. Relying on DMEP to help manage demand uncertainty, we did not place BASE order 4 in Q4 2008. As the downturn continued for another quarter (see Figure 8) and a new demand forecast was issued, we did not place BASE order 5 in Q1 2009 (see Figure 7). With DMEP, we simply held the FLEX options, waiting to determine whether the market would continue to decline, level off, or return to previous levels. If we had used the old method and taken the risk of the market coming back more quickly than our intuition of two years—that is, kept taking equipment deliveries and installing them, absorbing both the depreciation and the cost of keeping them warm—we would have realized revenue immediately if an unex- pected upturn occurred. However, this is equivalent to betting $500 million worth of specialized equip- ment on a market upturn and paying $22 million per month while waiting. We could have taken that same risk for the same potential reward by exercising the Figure 9: The graphic shows the recovery plan to quickly intercept the Q1 2009 demand forecast, including BASE orders 6 and 7 and FLEX orders 1 and 2. Without the FLEX orders to ll the capacity gaps left by the cancellation of BASE orders 4 and 5, Intel would not have realized the demand ful llment upside. options instead of holding them, but we decided not to do so. In Q2 2009 after two quarters of rapid decline (see Figure 8), the market gave indications of begin- ning a comeback, and capacity procurement decisions had to be made about BASE order 6 (see Figure 7).
Prior to implementing DMEP, betting on this upward trend would have meant (re)ordering equipment in Q2 2009 with a four-quarter lead time to intercept the demand forecast in Q2 2010 (see Figure 7). Capacity would have remained at in Q4 2009 and Q1 2010 because BASE orders 4 and 5 had not been placed.
With DMEP, we had the possibility to recover and we took it. As Figure 9 shows, in Q2 2009 we had the opportunity to place BASE order 6 and exercise FLEX options (FLEX order 1). Because the latter involved a two-quarter lead time, they served to substantially ll in the capacity gap left by the nonplacement of BASE order 4. As the market continued its upturn in Q3 2009 (see Figure 8), another recovery possibil- ity presented itself and we again took it. As Figure 9 shows, in Q3 2009 we had the opportunity to place BASE order 7 and exercise FLEX order 2. The latterBASE: order 8 Q1/08 forecast BASE: order 7 FLEX: order 2 BASE: order 5 BASE: order 6 FLEX: order 1 •• BASE: order 1 BASE: order 2 BASE: order 3 BASE: order 4 DEMAND FULFILLMENT UPSIDE Q4/08 reforecast Q1/09 reforecast Q1 – 08 Q2 – 08 Q3– 08 Q4– 08 Q1– 09 Q2– 09 Q3– 09 Q4– 09 Q1–10 Q2–10 Q3–10 Q4–10 Q1–11 Q2–11 Downloaded from informs.org by [132.174.254.97] on 25 February 2017, at 11:22 . For personal use only, all rights reserved.
Kempf et al.:
Optimizing Capital Investment Decisions at Intel Corporation Interfaces 43(1), pp. 62–78, © 2013 INFORMS 75served to substantially erase the capacity gap left by the nonplacement of BASE order 5. The bene t from the FLEX mode is in the demand captured by the capacity in the shaded area in Figure 9. This demand represents the capacity differ- ence between not having the FLEX mode (the lower capacity pro le, as Figure 7 shows) and exercising FLEX orders 1 and 2 (the higher capacity pro le, as Figure 9 shows). Intel's central nancial enter- prise services organization has validated the upside nancial bene t. Its calculations consisted of (the dif- ference in wafer start capacity per week with and without FLEX orders) times (the 26-week pull in using reduced lead-time options) times (the number of product chips per wafer) times (the average sell- ing price of the products in that period). Using this con dential data, the result was a $2.24 billion rev- enue upside. It is worth noting that during the peri- ods in which the revenue improvement was realized (Q4 2009 and Q4 2010), Intel's gross pro t margin increased from 58 percent to 67 percent.
DMEP has since been applied to the transition from the 32 nm manufacturing process to the subsequent 22 nm process. As Figure 1 shows, it is too soon to quote the bene ts. However, it is clear that the options with reduced lead time can be (1) exercised late in a normal market to manage uncertainty, reduc- ing cancellation costs and overpurchase of capacity; (2) exercised early to take advantage of a higher- than-expected demand and capture upside revenue; (3) held in a sluggish market to avoid unnecessary costs; or (4) cancelled in a market with much lower- than-expected demand.
Bene ts to the Capacity Supply Chain The simple yet effective framework has resulted in signi cant changes in the supply chain. For exam- ple, the historic industry problem of long lead-time lithography scanners has been mitigated. DMEP also changed Intel's relationships with its suppliers. Intel has actively sought to transform the source and procurement of capital equipment from an at-arms- length legal transaction to a collaborative partnership in which both sides bene t. The supplier enjoys con- dence in the higher probability that the equipment will be ordered eventually so that Intel can avoid option expiration costs. Similarly, because the equip- ment can be quickly turned from order to delivery, the option provides the supplier faster time to cash than a conventional tool sale. Overcoming the challenges and delivering these documented quantitative and qualitative results has been recognized at the highest levels inside Intel.
In August of 2011, Intel's CEO and his staff awarded the Intel quality award (IQA) to the TME organiza- tion. The velocity program with DMEP was a key component in determining the recipient of this award.
The IQA, which is Intel's highest award, recognizes internal organizations for their continuous improve- ment efforts. DMEP has also helped Intel ascend the annual Gartner top-25 supply chains, which comprise the best global performers as measured by nan- cial and supply chain performance metrics. Intel rose from 25th in 2009 to 7th in 2012. This represents the largest jump among all peer companies and is unprecedented, given that Intel is the only nondirect- to-the-consumer company on the list.
Insights and Concluding Remarks The capital supply chain velocity program with DMEP has transformed the way we make capital investment decisions at Intel. According to Intel's CEO Paul Otellini, making capital equipment in the highly volatile environment in which Intel operates “is a dif cult decision process, involving nance, manufacturing, supply chain, and operations. A team of leaders at Intel in collaboration with researchers at Stanford has met this challenge head-on through the application of OR. Their insights have led to land- mark advancements in optimizing capital investment decisions. The team has done nothing short of cre- ating a win for all stakeholders—for Intel, for our equipment suppliers, and for everyone whose lives are enriched across the compute continuum.” DMEP transformed the previous art of balancing billion-dollar revenues from dynamic uncertain mar- kets against billion-dollar spends on long lead-time equipment sets into a science. Our OR-based model supports informed decision making at the strategic, tactical, and operational levels. DMEP has provided Intel hundreds of millions of dollars in documented cost savings and at least $2 billion in revenue upsideDownloaded from informs.org by [132.174.254.97] on 25 February 2017, at 11:22 . For personal use only, all rights reserved.
Kempf et al.:
Optimizing Capital Investment Decisions at Intel Corporation 76 Interfaces 43(1), pp. 62–78, © 2013 INFORMSfor one manufacturing process transition during a period of economic crisis. More importantly, at every biennial manufacturing process transition, compared to our previous approach, Intel now has the oppor- tunity to avoid unnecessary cost, capture upside rev- enue, or both, while maintaining a high service level for our customers. DMEP has helped strengthen our working relation- ships with our suppliers. Our long-term collaboration on the equipment technology required to realize Moore's Law has been extended to shorter-term col- laboration on the win–win business arrangements necessary to better manage market uncertainty. Over time, we will be able to build a stronger semiconduc- tor ecosystem that will in turn bene t all customers in all semiconductor-related markets with more capable products at lower prices.
Delivering equipment at FLEX mode lead time was initially very challenging for our equipment sup- pliers. Strong collaboration was required to reengi- neer the supply chain to be capable of delivering these shorter lead times. As this reengineering has taken place, we have seen BASE mode lead times also decline. We expect this supply chain ef ciency improvement to continue, bene tting Intel, equip- ment suppliers, and the semiconductor ecosystem.
DMEP is not Intel speci c or limited to the semi- conductor industry. Any company that modi es its capacity based on forecasts of its market can ben- e t from applying the concepts and the OR-based approach. Obviously, the more uncertain the forecast, the longer the lead time of the capacity, and the more cost and revenue involved, the more bene cial DMEP will be. Industries such as automotive and aerospace come quickly to mind. However, any capital-intensive manufacturing rm with OR skills can implement DMEP and improve its bottom line. This project, like many others we have undertaken over the past 25 years at Intel, has con rmed that the math is necessary but not suf cient. On re ection at the end of any OR-based project, success depends on the people involved. This is true from many perspectives; this project exempli es most of them.
Collaboration on practice, theory, intuition, and ana- lytics was the foundation of our success. Early in the project, the collaborators were Intel TME and DE; they expanded later to include Stanford University. Prior collaboration between Stanford and Intel on other OR/MS problems allowed for a quick ramp-up on this problem. Deep specialized expertise mixed with an open mind characterized each member of the team.
The results speak loudly for themselves. The project also led to several theoretical advancements and in u- enced graduate-level courses at Stanford University by providing a deeper understanding of how theoret- ical OR/MS techniques can be used to provide valu- able decision support in practice. As has often been our experience, a successful project breeds related efforts. In this case, our collabo- ration on optimizing capital investment decisions has spawned additional projects in our capital equipment supply chain, including OR-based models to increase the speed at which we can build and ramp up a new factory and OR-inspired approaches to further insure supplier ecosystem health. Additional improvements in this area will further support our corporate goal of creating and extending computing technology to connect and enrich the lives of every person on earth.
Appendix. The Stochastic Program for the Execution Module The selling season of the product is Nperiods and starts in period 1. To prepare for the demand ramp, Intel starts placing orders L b periods before the rst demand realization and stops placing orders L f periods before the end of the selling season. Therefore, the length of the planning horizon is NCL bƒ L f periods. The rm orders only from the BASE mode in periods 1 ƒL b1 0 0 0 1 L f; it orders from both modes in periods 1 ƒL f1 0 0 0 1 N ƒL b; and it orders only from the FLEX mode in periods NƒL bC 11 0 0 0 1 N ƒL f. In periods Nƒ L f C 1 to N, Intel simply satis es demand with the existing capacity. Speci cally, in period mof the planning horizon ( m D1ƒ L b1 2 ƒ L b1 0 0 0 1 N ƒL f5 , the execution module solves the following stochastic program:
maximizeE B 12 N 3E F 12 N d1_ 4m C1510001d N N X i D 1„ i 8p is i ƒ c bB iƒ c f F i ƒ c h K i9 ƒ „N C1 c u d rem N C1 subject to 2For each realized sample path d1_ 4m C151 0 0 0 1 d N2 s i D min 8k i1 d iC drem i 9 for iD 11 0 0 0 1 N 1 k i –d iC drem i for iD 1_ 4m CL f51 0 0 0 1 N 1 k iD k iƒ 1 C B iC F i3 for iD 11 0 0 0 1 N withk 0 D 01Downloaded from informs.org by [132.174.254.97] on 25 February 2017, at 11:22 . For personal use only, all rights reserved.
Kempf et al.:
Optimizing Capital Investment Decisions at Intel Corporation Interfaces 43(1), pp. 62–78, © 2013 INFORMS 77d rem i D 4d iƒ 1 C drem i ƒ 1 ƒ k iƒ 15C 3 for iD 11 0 0 0 1 N withdrem 1 D 01 N X i D 1B i BT 3 N X i D 1F i FT 1 E B 12 N 03 E F 12 N 01 E B 12 m ƒ1C L b D E B 1 2 m ƒ1C L b3 E F 12 m ƒ1C L f D E F 1 2m ƒ1C L f 0 We de ne x_ y 2 Dmax 4x1 y5 and use the notation E x 12j to represent the vector [ x 11 x 21 0 0 0 1 x j7 T . Additionally, p i is the pro t margin for period i; c b the BASE mode execu- tion price; c f the FLEX mode execution price; c h the unit equipment holding cost; c u the unit penalty cost for unmet demand after the planning horizon ends; d i the incoming demand in period i; drem i the unsatis ed demand remain- ing from period iƒ 1; s i the actual sales quantity during period i; k i the cumulative capacity position at period i; – the service-level target; and „the discount factor. B i and F i represent the optimal decisions already executed in pre- vious periods. We note that the subscripts used for order quantities denote the period in which the ordered equip- ment arrives.
Acknowledgments The authors thank their Edelman competition coaches, Grace Lin and Atul Rangarajan, whose valuable sugges- tions improved the content and exposition of the paper and the corresponding presentation signi cantly. The authors express their deepest gratitude to Robert Bruck (Intel Cor- porate Vice President of Technology Manufacturing Engi- neering) and Blake Sacha (Intel Capital Planning Director) for valuable discussions and nancial support. The authors also thank Paul Otellini (CEO, Intel Corporation) and Andy Bryant (Chairman, Board of Directors, Intel Corporation) for their video testimonials supporting this work. Chen Peng's current employment at Google is not con- nected in any way with the Intel-funded work he did while a graduate student at Stanford University that is part of this Edelman Award application. Authors Feryal Erhun and Chen Peng were partially supported by National Science Foundation CAREER Award [NSF/CAREER-0547021].
References Cachon GP (2003) Supply chain coordination with contracts. De Kok AG, Graves SC, eds. Supply Chain Management: Design, Coordination and Operation (Elsevier, Amsterdam), 229–340.
Chao X, Chen H, Zheng S (2009) Dynamic capacity expansion for a service rm with capacity deterioration and supply uncer- tainty. Oper. Res. 57(1):82–93.
Daniel KH (1963) A delivery-lag inventory model with emergency. Scarf HE, Gilford DM, Shelly MW, eds. Multistage Inventory Models and Techniques (Stanford University Press, Stanford, CA), 32–46. Feng Q, Sethi SP, Yan H, Zhang H (2006) Are base-stock policies optimal in inventory problems with multiple delivery modes?
Oper. Res. 54(4):801–807.
Fukuda Y (1964) Optimal policies for the inventory problem with negotiable lead time. Management Sci.10(4):690–708.
Graves SC, Meal HC, Dasu S, Qui Y (1986) Two-stage production planning in a dynamic environment. Axsater S, Schneeweiss C, Silver E, eds. Multi-Stage Production Planning and Inventory Con- trol (Springer-Verlag, Berlin), 9–43.
Hausman WH (1969) Sequential decision problems: A model to exploit existing forecast. Management Sci.16(2):93–111.
Liu L, Parlar M, Zhu SX (2007) Pricing and lead time decisions in decentralized supply chains. Management Sci.53(5):713–725.
Moore G (1965) Cramming more components onto integrated cir- cuits. Electronics 38(8):114–117.
Peng C, Erhun F, Hertzler E, Kempf K (2012) Capacity planning in the semiconductor industry: Dual-mode procurement with options. Manufacturing Service Oper. Management 14(2):170–185.
Shapiro A (2008) Stochastic programming approach to optimization under uncertainty. Math. Programming 112(1):183–220.
SIA Historical Billing Reports (2012) Historical billings from 1976 to present. Accessed April 1, 2012, http://www.sia-online .org/industry-statistics/historical-billing-reports.
Vaidyanathan V, Metcalf D, Martin D (2005) Using capacity options to better enable factory ramps. Intel Tech. J.9(3):185–192.
Van Mieghem JA (2003) Capacity management, investment, and hedging: Review and recent developments. Manufacturing Ser- vice Oper. Management 5(4):269–302.
Yazlali Ö, Erhun F (2009) Dual-supply inventory problem with capacity limits on order sizes and unrestricted ordering costs.
IIE Trans. 41(8):716–729. Karl G. Kempf is an Intel Fellow and director of deci- sion engineering for the Intel Architecture Group at Intel Corporation. He focuses on product design and develop- ment decision problems and previously was responsible for manufacturing and supply chain decision problems. He has been involved in designing and implementing decision policies for production scheduling, equipment maintenance, factory ramp management, equipment selection and layout, strategic and tactical production planning, inventory plan- ning, demand forecasting, logistics operations and product design, as well as many modeling and simulation projects.
He is a member of the National Academy of Engineering and has published widely. He holds a BS in chemistry, a BA in physics, a PhD in applied mathematics, and he com- pleted postdoctoral studies in computer science.
Feryal Erhun is an assistant professor of management sci- ence and engineering at Stanford University. His research interests are in the strategic interactions between stakehold- ers in supply chains. His research on nonpro ts, health- care operations management (OM), and sustainable supply chains has informed the development of theory address- ing the unique challenges arising for these organizations and has extended traditional OM theory to these new and important settings.Downloaded from informs.org by [132.174.254.97] on 25 February 2017, at 11:22 . For personal use only, all rights reserved.
Kempf et al.:
Optimizing Capital Investment Decisions at Intel Corporation 78 Interfaces 43(1), pp. 62–78, © 2013 INFORMSErik F. Hertzler is the capital equipment supply chain velocity and exibility staff engineer in Intel’s Technology Manufacturing Engineering Capital Planning group. His current work includes the application of statistical model- ing to improve capital supply chain velocity, collaborative university research efforts to use model predictive controls to synchronize supply-demand forecasting, and application of lean methods to streamline existing capital business pro- cesses. He earned a bachelor’s degree in mechanical engi- neering from the University of New Mexico in Albuquerque and is a PhD candidate in the Arizona State University Industrial Engineering Department.
Timothy R. Rosenberg is a senior nance manager in Intel’s nance organization, serving as a senior nance manager in the Intel Architecture Development Group. He supports nancial analysis for IP planning strategies in Intel’s system-on-chip businesses; he also supports third- party IP vendor strategies. He holds a bachelor 's degree in nance from the University of Nebraska and an MBA in business administration with a concentration in nance from Arizona State University. Chen Peng earned a doctorate in the Department of Man- agement Science and Engineering at Stanford University. He works at Google as a quantitative analyst involved in cross- functional projects in forecasting and planning Google’s global data center capacity as well as the optimal allocation of Google’s cloud computing resource.Downloaded from informs.org by [132.174.254.97] on 25 February 2017, at 11:22 . For personal use only, all rights reserved.
