Soochow Journal of Economics and Business
No.70 (September 2010)：77-102.
Strategic Behavior in The Semiconductor Industry:
The Case of The DRAM Market
Daw Ma*
Abstract
The growth in demand for electronics components and computers in turn gives rise to
more and more demand for semiconductor devices, both in terms of quantity and innovation.
However, in view of the huge capital investment and technology required together with the
unique industry structure, the semiconductor manufacturers are faced with substantial risks.
Therefore, analyzing the market’s structure and the industry’s strategic behavior are crucial
in the semiconductor industry. In this paper, the author analyzes the strategic behavior of the
DRAM firms using an n-firm Stackelberg competition model and compares this with a generic
Cournot model. An early-mover advantage in DRAM production is suggested. Furthermore,
a strategy for DRAM manufacturers is to commit themselves to more capacity in order to
deter further entrants. Such behavior may also explain the volatility and the common overcapacity problem in the DRAM market.
Keywords: strategic behavior, semiconductor industry, DRAM, early mover advantage,
entry deterrent
*
Chung-Hua Institution for Economic Research75 Chang-Hsing St., Taipei 106, Taiwan, ROC
Tel: 02 27356006 ext 634 E-mail:[email protected]
第七十期
1. INTRODUCTION
The semiconductor industry is considered by many countries to be a strategic industry.
The semiconductor is the fundamental element of all modern electronic systems and computers, such as mobile phones, personal computers, consumer electronics, and network communications equipment. The growth in the demand for electronics has led to an increase in
the demand for semiconductor devices and the accompanying sustained innovation. Yet
with the consequent requirements for huge capital investment and research and development
into the rapidly changing technology, semiconductor manufacturers are faced with substantial risks. Thus the analyses of the market structure and the formation of manufacturing strategies are crucial in the semiconductor industry.
The semiconductor industry had total worldwide sales of $235 billion in 2009. From
1983 through 2009, the industry grew vigorously at a staggering average compound annual
growth rate (CAGR) of about 12 percent. Very few industries can boast such a growth rate
over a period as long as the past 27 years! This brings a strong long-term outlook and high
hopes for the industry. Nevertheless, if we look at the changes in year-to-year growth rates,
the short-term fluctuations make the industry difficult to operate. The changes in growth
rates are seen to swing from a positive 37 percent in 2000 to a negative 32 percent in the following year as shown in Figure 1. The short-term fluctuations are huge and were largely
contributed by the market supply and demand imbalance. Together, these inbuilt problematic characteristics have resulted in a vicious industry cycle.
Without a doubt, firms face a real dilemma when operating in this industry, and strategic behavior engaged in by semiconductor firms to reduce competition by both actual and
potential rivals is particularly worth investigating. To analyze the strategic behavior of the
industry, the Dynamic Random Access Memory (DRAM) market is selected as the target industry. First, the study identifies and analyzes the characteristics of the semiconductor industry and the DRAM market. Based on the identified characteristics, an industry-specific
DRAM competition model can then be carefully constructed. A game theoretic model is needed in order to capture the interdependence of these characteristics. In this paper, a generalised Stackelberg model is used to analyse the competitive behaviour of the DRAM firms.
半 導 體 產 業 的 策 略行 為 ： 以 DRAM 市 場 為例
The Stackelberg model, developed by Heinrich von Stackelberg in 1934, is probably
one of the most widely cited models of non-cooperative oligopoly behavior. In the Stackelberg model, one firm takes the other firms’ reactions as given and each of the firms
chooses outputs sequentially. Anderson and Engers (1992) and Boyer and Moreaux (1986)
provide solutions to the generalized Stackelberg model. Other researchers, such as Church
and Ware (1996), Robson (1990) and Vives (1988), incorporate sunk fixed costs and analyze
the Stackelberg model where the number of firms is determined endogenously. Pal and Sarkar (1998; 2001) analyze the n-firm Stackelberg model with non-identical firms. Sarmento
and Brandao (2007) study the way a multiproduct firm can develop a price strategy that use
the regulatory policy to deter entry under Stackelberg competition. Jiang, Hu and Chen
(2010) study the decision-making processes of knowledge transfer as the Stackelberg leader-followers games between the core firm and partners in the technology innovation alliance.
Furthermore in the present study, a comparison is made between the DRAM competition model and a generic one without the industry-specific characteristics. The analysis of
the comparison can shed the light on the strategic behaviors of the DRAM firms. A strategy
for DRAM manufacturers is to commit themselves to more capacity in order to deter further
entrants. The equilibrium number of firms is less than under Cournot competition. Anderson
and Engers (1992) also show that the n-firm Stackelberg price is lower and quantity is higher. Church and Ware (1996) share similar findings that firms in their sequential entry model
will select higher output to deter entry. In addition, a strong early-mover advantage in
DRAM production is also suggested in this paper. This finding also coincides with the “enter
early” strategy that Gruber (1992) suggested in the semiconductor industry.
第七十期
60%
46%
42%
40%
38%
37%
29%
28%
23% 23%
19%
8%
4%
8%
18%
10%
7%
4%
1%
-20%
3
5
7
200
200
200
5
199
1
3
199
4%
200
1
199
9
9
198
-9%
199
7
198
7
5
198
199
3
198
0%
9%
9
20%
-5%
200
20%
32%
-10%
-9%
-17%
-32%
-40%
source :IC Insights
Figure 1. Worldwide Semiconductor Industry Growth Rates
2. SEMICONDUCTOR MANUFACTURING AND
THE DRAM MARKET
A semiconductor belongs to a class of materials that can assume the properties of either
a conductor or an insulator. The most widely used semiconductor material is silicon. However, materials such as germanium and gallium arsenide are also common. Transistors amplify or switch electrical current and other circuit elements that are “integrated” on a single
piece of semiconductor material to form an integrated circuit (IC). Integrated circuits are the
building blocks of our modern electronic systems.
The process of manufacturing an integrated circuit involves building up a series of layers on a round slice of polycrystalline silicon wafer. Building up each layer involves a sequence of steps. A transistor consists of a gate and a channel, separated by a thin layer of insulation. Applying voltage to the gate determines whether or not current can flow through
the channel. Normally, the various parts of a transistor are made using photolithography.
半 導 體 產 業 的 策 略行 為 ： 以 DRAM 市 場 為例
This involves shining light through a patterned mask and on to the surface of a silicon wafer,
so as to change the chemical properties of the exposed parts of that wafer. Once this is done,
a series of chemical processes, some of which require high temperatures, are used to etch
microscopic electrical circuits on to the wafer’s surface. After the processing is completed,
dies or chips are cut from the silicon wafer.
The number of silicon chips that can be produced at one time affects the cost of making
them. The size of each batch is determined by the diameter of the silicon wafer on which the
hundreds of identical chips are etched. Bigger wafers mean more chips per batch—which,
in turn, mean lower processing costs per chip. The move from 200mm (eight-inch) to
300mm (twelve-inch) diameter wafers has provided 125% more chips than a 200mm one,
and has also reduced costs by a further 20%. Semiconductor manufacturing is essentially a
fixed-cost business with steep price competition over time. The new 300mm wafer fabrication plants, “fabs” ,cost well over $2 billion each.
Faced with the huge costs of building a new fab, one of the objectives of manufacturers
is to reach a “critical mass” of production as quickly as possible. There is a learning curve
to ramp up production. High asset utilization is necessary for profitable operations. As experience in production accumulates, these problems at the manufacturing stage can be minimized, and the proportion of satisfactory chips can be improved, hence the increase in the
yield rate. Defects on chips can also affect the yield. These are caused by impurities in the
silicon wafer or because of teething problems with a new chip design. Initial yields may be
as low as one-third to one-half of total production, but 90 percent or better may be attained
within a period of twelve to eighteen months (Sutton, 1998).
In chip making, size is everything—the smaller the better. Conventional transistors
typically have a gate length of 50 nm (nanometers, each one billionth of a meter.). Shrinkage
of gate length will lead to a reduction in chip geometry or die size. As die size decreases, the
number of chips per wafer increases and hence productivity rises. The turnover of the semiconductor technology is also very rapid. Moore’s law is an empirical, but remarkably prescient, observation perceived by Gordon Moore in 1965 who later co-founded Intel. Moore’
s law states that the number of transistors that can fit on a computer chip—and thus the capacity to crunch numbers—doubles every 18 months, thus allowing computers to double in
第七十期
speed, or to fall in price by half.
Dynamic Random Access Memory (DRAM) is a memory storage device, which is
used in electronics and primarily in personal computers. DRAM uses charge storage on a
capacitor as an equivalent to binary data and requires data to be continually regenerated.
DRAM can be randomly accessed at any memory cell, rather than having to proceed sequentially from a starting cell, and needs to have its storage cells refreshed or given a new electronic charge periodically. The periodic refreshing of cells renders this type of memory dynamic.
A single transistor DRAM cell was first invented in 1966 by Dr. Dennard at IBM. Virtually all modern DRAMs are based on the design of this single transistor cell. The first
commercial DRAM was introduced by Intel to the market in October 1970. This 1Kbits
DRAM, referred to as i1103, was manufactured as a 6 mask silicon gate PMOS process and
sold for around $21 (IC Knowledge, 2000). By the year 1971, more than 100,000 units of
DRAM were sold in commercial quantities. DRAMs are used as the main memory in most
memory systems and have the highest volume production. The global DRAM market was
valued at $21.9 billion in 2009. DRAMs account for a major portion of the semiconductor
market. The main application for DRAMs is the personal computer. About 80 percent of
DRAM were used in computers (IC Insights, 2007). In a PC, DRAM serves as the main
memory and holds all of the “active” information that the computer uses. Beyond the PC
market, applications of DRAM are divided among several end-uses such as graphics, communications, and consumer products.
There are a few variations of the DRAMs which may be categorized in terms of different architectures. All the DRAM architectures strive to address one fundamental issue facing the traditional DRAMs—the rapid transfer of a significant amount of memory to wide
bandwidth, high-performance microprocessors. Examples of the DRAM categories include
Synchronous DRAM (SDRAM), double data rate (DDR) DRAM, Rambus DRAM
(RDRAM), video RAM (VRAM), synchronous graphics DRAM (SGRAM), and so on. The
manufacturing processes and technology are becoming more sophisticated along with the
progress in the product designs and generations. Each generation is classified by capacity.
To date, the DRAM generations include 1Mbit, 4Mbit, 16Mbit, 64Mbit, 128Mbit, 256Mbit,
半 導 體 產 業 的 策 略行 為 ： 以 DRAM 市 場 為例
1Gbit and 2Gbit.
The DRAM market is very volatile and highly cyclical in nature. The DRAM industry
is characterized by a high degree of product homogeneity and a fairly high level of setup
cost and R&D. The DRAM market is quite concentrated. As shown in Table 1, the 2009
DRAM market was dominated by Samsung, Hynix, and Elpida. Together, these three firms
accounted for more than 72.7 percent of the DRAM market in 2009.
Table 1. 2009 DRAM Market Shares
Rank
Company
2009Q4
Sale(M USD)
2009Q4
Market Share
1
Samsung
2,750
31.7%
2
Hynix
1,871
21.6%
3
Elpida
1,680
19.4%
4
Micron
1,057
12.2%
5
Nanya
496
5.7%
6
Powerchip
408
4.7%
7
Winbond
166
1.9%
8
ProMOS
86
1.0%
Others
168
2.0%
Total
8,682
100.0%
Source: DRAMeXchange
3. DRAM COMPETITION MODEL
In this study, we employ an n-firm sequential Stackelberg oligopoly to model the
DRAM manufacturing competition. The defining characteristic of the competing DRAM
manufacturers is the interdependence of their strategies in a sequential manner as Stackelberg suggested. Given the sheer size of the market and the high concentration, any prospective entrant with additional capacity will have a significant impact on the DRAM market
structure. Conversely, given the vast size of investment and market size, the entry decision
of any potential manufacturer is going to heavily depend on the output levels of established
manufacturers. Therefore it is crucial for any player in the market to observe the market sup-
第七十期
ply and all individual firms’ outputs as precisely as possible, especially where there are only
a few firms. The quantity supplied by a DRAM manufacturer can be normally observed by
the size of its production plant (200mm or 300mm fab), the level of the adopted technology
(70, 50 nm, and so on), and the degree of manufacturing learning such as the yield rate1. If
all firms observe the outputs of the others and observe them well, we assume that the capacity choices of the DRAM manufacturers will become common knowledge. Firms can be assumed to enter in sequence because ‘some entrants become aware of a profitable market before others or require longer periods of time in which to “tool up”’(Prescott and Visscher,
1977). In the case of semiconductor production, immense capital investment and a lead time
of two years is required to establish a manufacturing facility. The early entrants are often
those wise ones who spotted the potential profitability and committed a vast capital investment at an early stage. However, while the successes of the early entrants induce further entries, the later entrants have to wait in line to tool up their production plants.
We also assume all firms are identical in the sense that the cost function of each firm
is approximately the same. The fixed costs of setting up and running a production facility
are more or less uniform regardless of the locations. Machinery and equipment accounts for
85 per cent of the total set up costs of some $2 -3 billion (IC Insights, 2002). The impact that
this has on asset management is further illustrated by the fact that the rate of technological
obsolescence is highest on the equipment side, with most new equipment typically only lasting two device generations. Depreciation of the facility capital cost is the dominant aspect
of cost. This depreciation accounts for almost 70 percent of the processed wafer cost.
半 導 體 產 業 的 策 略行 為 ： 以 DRAM 市 場 為例
Probe
Yield Loss
Chemicals/Gases
Masks
Maintenance
Utilities
Wafer
Other
Labour
(Total Processed Wafer Cost = $1,400)
Depreciation
$0
$100
$200
$300
$400
$500
$600
$700
$800
$900
$1,000
Source: Future Horizons
Figure 2. 200m Processed Wafer Cost Breakdown
Apart from depreciation, virtually all of the other costs are also fixed, except for the
wafer itself and the small amount of chemicals, speciality gases and other consumables.
Modern wafer fabs are highly automated facilities, requiring a highly-skilled team of technicians to keep the factory up and running. The cost difference of this highly-skilled labour
in different parts of the world is considered marginal since depreciation is the dominant cost
of production. For the reason of the sizeable set up costs spent on machinery and equipment
and the high percentage of depreciation as a proportion of total production costs, our assumption of identical firms is consequently justified.
We assume that there are n firms in the semiconductor market. Each firm i i ＝ 1, ..., n
produces a homogeneous product-Dynamic Random Access Memory (DRAM). We assume
complete information which means that each firm knows the DRAM market demand and
cost functions of all firms. Let the quantity produced by firm i be q i q i ＞ 0 i and
qi
0, L i where L i, ..., L n denotes the capacity limits of the firm. Let the aggregate supply
n
produced in the DRAM market be Q＝
i＝1
q i and let the inverse demand function be
P Q ＝ a bQ, where a ＞ 0 and b ＞ 0. We assume that all firms’ technologies are symmetric. We assume that firm i has a cost function c i q i ＝ F ＋ cq i where c is a constant
第七十期
c
0 and F is the fixed setup cost for all firms.
Under the setting which we described, DRAM manufacturers sequentially decide how
much to produce. This sequential entry game consists of n stages. In stage i i ＝ 1, ..., n ,
firm i decides whether or not to enter the market and if it enters, it commits itself to an output
level. The solution concept adopted in this study is subgame perfect Nash equilibrium, and
a game of this kind can be normally solved by backward induction. Two conditions must be
satisfied in Nash equilibrium. First, for the entrant, the final stage profits must prove to be
viable, otherwise the firm will not choose to enter the market
i
0 . The second condition
states that there will be no gap in the equilibrium market after all n firms have entered at the
end of the game. Any marginal firm n ＋ 1 will not have positive payoff given the same
market size. To play the game, firm 1 simply selects an output first. To follow, firm i i 1
produces after the firm which precedes it i 1 . The respective quantity choices of firms 1
to i 1 are taken as given when firm i selects its own output. During the game, firm i is perfectly aware of the number of followers n 1 prepared to enter the DRAM market. At the
same time, it is also acknowledged that the output choice of firm i will have an impact on
the quantity choices of its followers i ＋ 1, ..., n .
In order to select the optimal output, the strategy for firm i is a reaction function
i −1
Ri (∑ q j ) which assigns a level of production according to the quantity choices in all prej =1
i
ceding firms. The total output in firms 1 to i is
q i and the sum of the followers’ reaction
1
n
functions in firms from i ＋ 1 to n is
k −1
∑ R (∑ q )
k =i +1
i≠n
k
j =1
j
.
In the game of sequential entry with perfect information, the strategy of firm 1 is simply
to choose a level of output. The respective strategies for each of firms 2 through n are the
best reactions of quantity supplied given the output choices of the preceding firms. The output selection of firm i, q i , depends on the output selection of all preceding firms.
半 導 體 產 業 的 策 略行 為 ： 以 DRAM 市 場 為例
i −1
i ≠1
Ri : ∑ q k → qi
k =1
q i ∈ [0, Li ] ∀i
S i ＝ q i is a strategy set for firm i to select its outputs. A subgame perfect equilibrium to this
game is a set of strategies, S *, where each firm’s objective is to maximize its profit. Once
firm 1 select its optimal output, the followers (firm i, i 1) will make their own profit-maximizing output decisions based on the quantity selections of all preceding firms.
S * = ((q1 , R2 (q1 ), R3 (q1 + q2 ),..., Rn (q1 + ... + qn −1 ) ) | Max π i , ∀i ), i = 1,..., n
The optimization of profit for firm 1’s output choice must satisfy the following condition:
n
k −1
n
k −1
⎤
⎡
q1 = arg max : π 1 (q1 , ∑ Rk (∑ q j )) = P ⎢q1 + ∑ Rk (∑ q j )⎥ q1 − c1 (q1 )
k =2
j =1
k =2
j =1
⎦
⎣
To select an optimal output level, it is required that firm i maximize its profit.
This can also be expressed as:
⎡ i
⎤
n
k −1
⎢
Max π i = P ∑ qi + ∑ Rk (∑ q j )⎥ qi − ci (qi )
⎢1
⎥
qi ≥ 0
k = i +1
j =1
i≠n
⎣⎢
⎦⎥
Using backward induction and similar procedures of Pal and Sarkar (1998), we first work
out the profit of the last firm n.
n
n −1
⎧
⎫
⎧
⎫
π n = q n ⎨(a − c) − b∑ q j ⎬ − F = q n ⎨(a − c) − bq n − b∑ q j ⎬ − F
j =1
j =1
⎩
⎩
⎭
⎭
第七十期
n 1
The first order condition of the above equation is a c
2bq n b
j＝1
q j ＝ 0 and we can de-
rive q n and Rn as
n −1
⎫
1 ⎧
qn =
⎨(a − c) − b∑ q j ⎬ = Rn
2b ⎩
j =1
⎭
Equation 7 represents the strategic decision of firm n after considering the output decisions
of all the preceding firms n 1 which in terms represents the reaction function of firm n
after received all the vital strategic information. Even firm n 1 selects its output prior firm
n, it still has to consider the strategic reaction of firm n in order to make the optimal strategic
decision of its own. That why an backward induction process is required in the analysis. The
profit function of firm n 1
n
1 is
n −1
n −1
⎧
⎫
⎧
⎫
π n −1 = qn −1 ⎨(a − c) − b(∑ q j + Rn ) ⎬ − F = qn −1 ⎨(a − c) − bRn − b∑ q j ⎬ − F
j =1
j =1
⎩
⎭
⎩
⎭
n −1
⎫
1 ⎧
(
)
a
−
c
−
b
q j ⎬ and hence
We then substitute R n of Equation 8 into
⎨
∑
2b ⎩
j =1
⎭
n 1
is now
n −1
n −1
n−2
⎡
⎫⎤
⎧
⎫
1⎧
1
π n −1 = qn −1 ⎢(a − c) − b∑ q j − ⎨(a − c) − b∑ q j ⎬⎥ − F = qn −1 ⎨(a − c) − bqn −1 − b∑ q j ⎬ − F
2⎩
2
j =1
j =1
j =1
⎭⎦⎥
⎩
⎭
⎣⎢
n−2
⎫
1⎧
(
a
−
c
)
−
2
bq
−
b
qj⎬ = 0
The first order condition of the above equation is therefore ⎨
∑
n −1
2⎩
j =1
⎭
and then
qn −1 =
n−2
⎫
1 ⎧
⎨(a − c) − b∑ q j ⎬ = Rn −1
2b ⎩
j =1
⎭
By continuing and observing the process of Equations 6-10, we obtain
⎡1⎤
πi = ⎢ ⎥
⎣2⎦
n −i
i
⎫
⎧
(
a
c
)
b
q i ⎬q i − F
−
−
⎨
∑
1
⎭
⎩
半 導 體 產 業 的 策 略行 為 ： 以 DRAM 市 場 為例
n−2
2
⎧
⎫
1
⎨(a − c) − b∑ q j ⎬q 2 − F , FOC this, we obtain q2 = {(a − c) − bq1 } .
2b
j =1
⎩
⎭
n −1
(a − c)
*
⎡1⎤
.
For π 1 = ⎢ ⎥ {(a − c) − bq1} q1 − F , FOC, we obtain q1 =
2b
⎣2⎦
⎡1⎤
For π 2 = ⎢ ⎥
⎣2⎦
2
(a − c) ⎡ 1 ⎤
. Moving this process forward, we have
Therefore q =
b ⎢⎣ 2 ⎥⎦
*
2
qi* =
(a − c) ⎡ 1 ⎤
b ⎢⎣ 2 ⎥⎦
i
i
n
(a − c) n ⎡ 1 ⎤
(a − c) ⎡ ⎡ 1 ⎤ ⎤ (a − c) ⎡
1⎤
=
Q = ∑q =
1− n ⎥
⎢1 − ⎢ ⎥ ⎥ =
∑
⎢
⎥
⎢
b i =1 ⎣ 2 ⎦
b ⎣⎢ ⎣ 2 ⎦ ⎦⎥
b ⎣ 2 ⎦
i =1
*
n
*
i
P* = a − bQ* =
a ⎡
1⎤
a + c(2n − 1)
+
−
=
1
c
2n ⎢⎣ 2n ⎥⎦
2n
π i* = qi* × P* − ci =
(a − c) 2
−F
2i + n b
4. STRATEGIC BEHAVIOR OF
DRAM MANUFACTURERS
In the previous section, the proposed DRAM competition model was established to
capture the characteristics of sequential entry and the interdependence of firms’ output selections in the DRAM market. Here, a generic Cournot model is employed to assume the relaxation of the above DRAM market characteristics. The behavior of the firm under the generalized Cournot model is analyzed and compared with the behavior under the DRAM competition model.
Recall the DRAM market inverse demand function P Q ＝ a b Q and firm i’s cost funcn
⎧
⎫
tion C i q i ＝ F ＋ cq i . We then calculate firm i’s payoff: π i = q i ⎨a − b∑ q j − c ⎬ − F and
j =1
⎩
⎭
its first-order condition
第七十期
n
a − b∑ q j − c − bqi = 0
j =1
Equation 16 represents the quantity selection of firm i under profit maximization. We sum
the above equation from firm 1 to n which results in
n
n
j =1
j =1
an − bn∑ q j − cn − b∑ q j = 0
n
This may be simplified to (a − c)n − (1 + n)b∑ q j = 0 then substituted in
j =1
n
a − b∑ q j − c − bqi = 0 to yield (a − c) −
j =1
(a − c)n
− bqi = 0 . Hence we obtain the Cournot
(1 + n)
quantity for firm i .
qiC =
(a − c) 1
b 1+ n
Then we can derive the Cournot aggregate quantity,
QC =
(a − c) n
b 1+ n
the Cournot market price
P C = a − bQ C =
a + cn
1+ n
and firm i’s Cournot profit
π iC = q iC × P C − c i =
(a − c) 2
−F
(1 + n) 2 b
The equilibria for both models are shown in Table 2.
半 導 體 產 業 的 策 略行 為 ： 以 DRAM 市 場 為例
Table 2. Equilibria for the DRAM Competition Model and Cournot Model
Equilibria for the DRAM Competition Model and Cournot Model
i
(a − c) ⎡ 1 ⎤
q =
b ⎢⎣ 2 ⎥⎦
(a − c) 1
qiC =
b 1+ n
*
i
Q* =
(a − c) ⎡
1⎤
1− n ⎥
⎢
b ⎣ 2 ⎦
QC =
P* =
(a − c) n
b 1+ n
a + c(2 n − 1)
2n
PC =
a + cn
1+ n
(a − c) 2
−F
2 i+n b
(a − c) 2
C
πi =
−F
(1 + n) 2 b
π i* =
Here, we also constructs examples of comparisons when the total numbers of firms in
the market are 2, 3, and 5.
Table 3. Two Firm Comparisons
n＝2
i＝1
i＝2
Stackelberg
a c 3
Q* ＝
b 4
a ＋ 3c
P* ＝
4
a
c 1
q *1 ＝
b 2
a c2
*
F
1＝
8b
a c 1
q *2 ＝
b 4
a c2
*
F
2＝
16b
Cournot
a c 2
QC＝
b 3
a ＋ 2c
PC＝
3
a
c 1
q C1 ＝
b 3
a c2
C
F
1＝
9b
2
a c 1
q C2 ＝
b
3
a c2
C
F
2＝
9b
Comments
Q* is higher
If a ＞ c and c only is a fraction of a,
thenP* is lower
q *1 is higher than q C1
*
1
＞
C
1
q *2 ＜ q C2
*
2
＞
C
2
Table 4. Three Firm Comparison
n＝3
i＝1
i＝2
i＝3
Stackelberg
a c 7
Q* ＝
b 8
a ＋ 7c
*
P＝
8
a c 1
*
q1＝
b 2
a c2
*
F
1＝
16b
a c 1
q *2 ＝
b 4
a c2
*
F
2＝
32b
a c 1
q *3 ＝
b 8
a c2
*
F
3＝
64b
Cournot
a c 3
QC＝
b 4
a ＋ 3c
C
P ＝
4
a c 1
C
q1 ＝
b 4
a c2
C
F
1＝
16b
a c 1
q C2 ＝
b 4
a c2
C
F
2＝
16b
a c 1
q C3 ＝
b 4
a c2
C
F
3＝
16b
Comments
Q* is higher
P* is lower
q *1 ＞ Q C1
*
1
＞
C
1
q *2 ＜ q C2
*
2
＞
C
2
q *3 ＜ q C3
*
3
＞
C
3
第七十期
Table 5. Five Firm Comparison
n＝5
i＝1
i＝2
i＝3
i＝4
i＝5
Stackelberg
a c 31
Q* ＝
b 32
a ＋ 31c
*
P＝
32
a c 1
*
q1＝
b 2
a c2
*
F
1＝
64b
a c 1
q *2 ＝
b 4
a
c2
*
F
2＝
128b
a c 1
q *3 ＝
b 8
a c2
*
F
3＝
256b
a c 1
q *4 ＝
b 16
a c2
*
F
4＝
512b
a c 1
q *5 ＝
b 32
a c2
*
F
5＝
1024b
Cournot
a c 5
QC＝
b 6
a ＋ 5c
C
P ＝
6
a c 1
C
q1 ＝
b 6
a c2
C
F
1＝
36b
a c 1
q C2 ＝
b 6
a
c2
C
F
2＝
36b
a c 1
q C3 ＝
b 6
a c2
C
F
3＝
36b
a c 1
q C4 ＝
b 6
a c2
C
F
4＝
36b
a c 1
q C5 ＝
b 6
a c2
C
F
5＝
36b
Comments
Q* ＞ Q C
P* ＜ P C
q *1 ＞ Q C1
*
1
＞
C
1
q *2 ＜ q C2
*
2
＞
C
2
q *3 ＜ q C3
*
3
＞
C
3
q *4 ＜ q C4
*
4
＞
C
4
q *5 ＜ q C5
*
5
＞
C
5
4.1 Entry Deterrent
By observing Table 2, when we compare the individual quantity produced by each firm,
q *i and q Ci, the DRAM leaders (the first few incumbents) tend to select higher outputs. We
also note that
This can also be clearly demonstrated using numeric values when n equals to 2, 3 and
5. The followers
i
2 in the DRAM market yield lower profits than under the Cournot
condition. The poor profitability of the DRAM followers is for the reason that the leaders
select higher outputs in order to induce subsequent entrants to cut back. Such actions of the
DRAM market leaders may be interpreted as entry deterrence. To better illustrate, if we say
the last firm to enter the market is x under the Stackelberg condition whereas y represents
the last firm to enter under the Cournot condition. To make the entry viable, firm x and y
半 導 體 產 業 的 策 略行 為 ： 以 DRAM 市 場 為例
must satisfy where
*
x
0 and
C
y
0. From Equation 15 and 21, we can show that. x ＜ y
The total number of firms is less under the Stackelberg condition. Therefore, by committing
more capacity than they would in the Cournot model, this is a strategy of the DRAM incumbents to discourage further competition. By observing Q* and Q C, we first notice that equilibrium output (Q*) is higher than its Cournot counterpart when n ＞ 1 as shown in Figure 3.
Q* =
a−c
1
a−c n
(1 − n ) > Q C =
, ∀n > 1
b
2
b 1+ n
This property of higher equilibrium aggregate outputs can support the over-capacity
phenomenon in the DRAM market. According to IC Insights (2002), over-capacity is one of
the major triggers which caused the downturn of the semiconductor industry cycles. We also
notice that, as shown in Figure 4, the market equilibrium price is lower.
a + c(2n − 1)
a + cn
P =
< PC =
, ∀n > 1
n
2
1+ n
*
The strategic behavior of the entry deterrent caused by over-capacity together with the
lower market prices will result in a defining characteristic in the DRAM market. The best
illustration can be found in the recent history of the DRAM market from 2006 to 2009.
第七十期
Note: a=200 b=1 c=0
Figure 3. DRAM Competition Model Quantity and Cournot Quantity
半 導 體 產 業 的 策 略行 為 ： 以 DRAM 市 場 為例
Note: a=200 b=1 c =0
Figure 4. DRAM Competition Model Price and Cournot Price
In 2006, the DRAM market had one of the best years in the semiconductor industry.
When the DRAM market took off in 2006, the DRAM maufacturers started to commit themselves to larger production capacity. The capital expenduture of DRAM manufacturers surged from $10.2 billion in 2005 to $15.6 billion and $18.7 billion in 2006 and 2007, respectively. It is a distinctive characteristic of the DRAM business for the market leaders to select
much higher outputs in order to induce subsequent entrants and/or incumbents to cut back.
The consequences of such strategic behavior can be found in Figure 5 where unit sales of
DRAM surged together with the even sharper fall in the average selling price. This deterrence strategy seemed to pay off in 2009. Not only were there no signs of new entrants, but
also one of the major players, Qimonda, was forced out of business. Serveral DRAM producers also struggled to survive and had to consider further consolidations or mergers.
第七十期
Source: IC Insights
Figure 5. Average Selling Price and Unit Sales of DRAM from 2006 to 2009
4.2 Early Mover Advantage
Since the market equilibrium price P* is lower and total outputs Q* are higher, the
higher output is more than offset by the lower price, and total profits are therefore lower in
the DRAM market. It is also noted that
*
1
＞
*
2
＞
*
3
＞ ... ＞
*
n 2
＞
*
n 1
＞
*
n
in the equilib-
rium DRAM competition model. The earlier the firm enters the market, the greater the profits the firm will make. Therefore, the early mover advantage is suggested in which an early
incumbent enjoys this kind of strong strategic asymmetry relative to its rival. There is clearly a financial incentive to become the leader in DRAM production.
This finding also coincides with the “enter early” strategy that Gruber (1992) suggested
in the Erasable Programmable Read Only Memory (EPROM)2 market. Flaherty (1992)
found that the empirical evidence in his study also indicates that the capacity development
rate had a much greater effect for firms that were ‘early movers’. By studying the learning
effect, Gruber (1998) suggested that intergenerational learning is pervasive, which seems to
provide competitive advantages to first movers. Steinmueller (1992) also stated that
learning provides the incentive to be among the first movers in the ‘new’ product generation.
半 導 體 產 業 的 策 略行 為 ： 以 DRAM 市 場 為例
Cost reduction through learning serves to lower the initially high prices that discourage the
broad utilization of new IC products. Capacity races are the consequence of attempting to
achieve higher profit margins and first mover advantages by getting started at the earliest
possible date with the ‘next’ generation of IC products.
Note: a=200 b =1 c =0 F =30
Figure 6. Profit of Firm i
第七十期
5. CONCLUSION
The semiconductor industry is highly competitive both in terms of price and technology. However, due to the huge capital investment and technology requirements together
with a unique industrial structure, the semiconductor manufacturers are faced with substantial risks. To analyze the strategies in semiconductor manufacturing, this study focuses on
conscious behavior arising among manufacturers in an oligopolistic DRAM market. The
early-mover advantage is suggested in DRAM production. There is a financial incentive to
become the leader. The DRAM market equilibrium price is lower and output and capacity
surplus are higher, and so total profits are lower. This is a strategy whereby DRAM incumbents can deter further entrants by committing themselves to higher quantities of output.
This over-supply behavior explains the characteristic of volatility in the DRAM market.
半 導 體 產 業 的 策 略行 為 ： 以 DRAM 市 場 為例
Footnotes
The yield rate is the percentage of wafers, dies or packaged units leaving a process as compared to
the amount of products that entered that process.
A non-volatile memory chip (memory is retained when the power is turned off) with the capability
to selectively erase information through a special electrical stimulus.
第七十期
REFERENCES
Anderson S. P. and M. Engers (1992), “Stackelberg Versus Cournot Oligopoly Equilibrum.” International Journal of Industrial Organization, 10, Iss.1, pp.127-135.
Boyer M. and M. Moreaux (1986), “Perfect Competition as the Limit of a Hierarchical Market
Game.” Economics Letters, 22, pp.115-118.
Church J. and R. Ware (1996), “Delegation, Market Share and the Limit Price in Sequential Entry
Models.” International Journal of Industrial Organization, 14, Iss. 5, pp. 575-609.
Flaherty, M. T. (1992), “Manufacturing and Firm Performance in Technology-Intensive Industries:
U S and Japanese DRAM Experience.” Review of Industrial Organization, 7, No.3-4, pp.273-294.
Gruber, H. (1992), “Persistence of Leadership in Product Innovation.” Journal of Industrial Economics, 40, No. 4, pp.359-375.
Gruber, H. (1998), “Learning by Doing and Spillovers: Further Evidence for the Semiconductor Industry.” Review of Industrial Organization, 13, No.6, pp.697-711.
IC Insights (2002), The McCLEAN REPORT 2002 Edition.
IC Insights (2007), The McCLEAN REPORT 2007 Edition.
IC Knowledge (2000), “History of the Integrated Circuit”, <Available from: http://www.icknowledge.
com/history/history.html>.
Jiang Z., L. Hu, and K. Chen. (2010). “Decisions of Knowledge Transfer in Technology Innovation
Alliance: a Stackelberg Leader-followers Model”. Operational Research, 10, No. 2, pp.231-242.
Pal D. and J. Sarkar (1998), “A Geometric Solution to a Stackelberg Oligopoly with Non-identical
Firms.” University of Cincinnati Working Paper.
Pal D. and J. Sarkar (2001), “A Stackelberg Oligopoly with Non-identical Firms.” Bulletin of Economic Research, 53, Iss.2, pp.127-134.
Prescott E. C. and M. Visscher (1977), “Sequential Location Among Firms with Foresight.” Bell
Journal of Economics, 8, Iss.2, pp.378-393.
Robson, A. J. (1990), “Stackelberg and Marshall.” American Economic Review, 80, Iss.1, pp.69-82.
Sarmento P. and A. Brandão (2007), “Entry Deterrence and Entry Accommodation Strategies of a
Multiproduct Firm Regulated with Dynamic Price Cap.”, International Advances in Economic Research, 13, No.1,pp.19-34.
Steinmueller, W. E. (1992), “The Economics of Flexible Integrated Circuit Manufacturing Technology.” Review of Industrial Organization, 7, No.3-4, pp.327-349.
半 導 體 產 業 的 策 略行 為 ： 以 DRAM 市 場 為例
Sutton, J. (1998), “Technology and Market Structure: Theory and History.” Cambridge: The MIT
Press.
Vives, X. (1988), “Sequential Entry, Industry Structure and Welfare.” European Economic Review,
32, Iss.8, pp.1671-1687.
第七十期
第七十期
（ 民 國 九 十 九 年 九 月 ）：77-102.
半導體產業的策略行為：以 DRAM 市 場 為 例
馬
道*
摘 要
半導體的數量與研發創新上的需求隨著對電子零件組與電腦需求的成長帶
動而逐漸增加，然而，鑒於龐大資本與技術的投入及獨特的產業結構，半導體
製造業者也面臨相當程度的風險。因此，分析市場結構和產業策略行為對半導
體產業將至關重要。本文作者運用 n 家廠商的 Stackelberg 競爭模型分析 DRAM
廠商的策略行為並且與一般的 Cournot 模型做對照，認為在 DRAM 市場中先行
者優勢是存在的；此外，DRAM 製造商致力提升於生產量以防堵其他的進入者
的策略可解釋在 DRAM 市場中常見的產能過剩與市場波動。
關鍵詞：策略行為，半導體產業，DRAM，早期進入優勢，嚇阻進入
*
中 華經 濟 研 究 院
Tel:0227356006ext634 E-mail: [email protected]