Competition and Quality Choice in the CPU Market
∗
Chris Nosko
Harvard University
November 2010
Abstract
This paper uses the CPU market to study how multiproduct firms generate returns
from innovation. Using a new dataset, I estimate a discrete-choice model of CPU
demand and then recover estimates of the sunk cost of product introductions. I combine
these estimates with a model of firm product choice to examine how product line
decisions change with asymmetric technological capabilities and with the competitive
environment. I use the model to show how technological leaders can use product lines as
strategic weapons, isolating competition to less desirable areas of the product spectrum.
I apply this insight to a large shift in technological leadership – Intel’s introduction of
the Core 2 Duo – and quantify the portion of returns that came from Intel’s ability to
push its principle competitor, AMD, into lower-margin product segments. I find that
competition plays a key role in determining firms’ product line decisions and that these
decisions are important in generating returns from innovation. Ignoring endogenous
product choices leads to underestimates of the social welfare losses from monopoly.
∗
I am grateful to Ulrich Doraszelski, Lisa Kahn, Greg Lewis, Julie Mortimer, Ariel Pakes, and Alan
Sorensen for invaluable advice. Also, I thank Brett Gordon for his help in putting together the pricing
data, Che-Lin Su for helpful discussions about computational methods, and especially Jeremy Davies of
Contextworld for generously supplying downstream data. All errors are solely mine.
1
1 Introduction
When firms innovate, they often don’t just introduce products at the top-end of the market.
Instead, they tend to reset their whole product line in an effort to extract the most possible
profit from their innovation. Their incentives to reshape lower market segments depend
on the industry structure, especially whether it is a monopoly or an oligopoly, and the
technological capabilities of rival firms. In oligopoly, one way that firms generate profit is

competitor can steal marketshare by introducing products at lower price points. This process
leads to a competitive equilibrium with more products spread throughout the product line.
Thus, in markets like CPUs, quality-based product separation can be driven largely by
competitive interaction rather than a desire to discriminate between consumer types.
2
This
contrasts with the standard literature on price discrimination, which sees the introduction
of quality-differentiated products as a mechanism for extracting more revenue from high-
valued consumers.
3
This finding implies that, because product line decisions matter little
for a monopolist’s profitability, when they innovate, resetting a product line will play much
less of a role in extracting profit from that innovation than for firms in an oligopoly.
I next find that, in oligopoly, returns from innovation come not only from the ability
to produce a better product at the top end, but also from an innovator’s ability to steal
business from rivals throughout the product line. Using a simple model, I construct an
example showing how a technological leader can isolate competition to lower margin portions
of the market, thereby increasing market power over a larger product space. I combine my
model with data from the introduction of the Core 2 Duo to quantify that role that these
business stealing effects played in generating profit for Intel. I break apart the portion of
returns that came from Intel’s introduction of new top products from the returns that came
from strategic quality choices throughout the product spectrum. This comparison gives
an estimate of how much we would underestimate the effect of competition on innovation
incentives if we held product lines fixed. Next, I compare the profits that Intel generated in
2
In a series of theory papers, Johnson and Myatt (2003, 2006) discuss the role of competition in driving
quality choices in a Cournot setting with differentiated products.
3
There is a long literature on using quality as a price discrimination mechanism going back at least to Jules
Dupuit in the 19th century. The foundational modern work is Mussa and Rosen (1978), with generalizations

In this paper I treat innovation outcomes as exogenous events, focusing on how firms generate returns
from them. With respect to the Core 2 Duo in particular, this is probably not a bad assumption: The project
that led to the Core 2 Duo, codenamed Banias, began at least as early as 2001 to develop a CPU for laptops
(which later became the Pentium M). It was only later that this technology was thought appropriate for the
desktop market. See the Seattle Times, “How Israel saved Intel”, More
generally, the results in this paper can be seen as an input to a dynamic game that endogenizes innovation
decisions.
4
estimation proceeds following the Pure Characteristic Model of Berry and Pakes (2007) with
some modifications tailoring the problem to my setting. I recover sunk costs of product
introductions from observing decisions on whether and when firms introduced new products
into the market. Moment inequalities (Pakes, Porter, Ho, and Ishii (2006) ) allow me to
recover bounds on the sunk cost parameter.
I find that a counterfactual monopolist has little incentive to introduce a whole product
line: With a single product he can capture 98% of the profit that he earns with a full,
optimally-placed product line. Given the sunk cost estimates, a monopolist would introduce
between 1 and 3 products compared to the 8 to 10 products that exist in the competitive
market. Consumer surplus from a monopolist is found to go down by 65% compared to the
competitive outcome. Much of that comes from the increased monopolist prices, but a non-
trivial 13% comes from the reduction of products and their inefficiently high quality levels.
I further find that the returns to innovation are higher in percentage terms in an oligopoly
than in a monopoly. My estimates indicate that Intel’s profits increased by 96% with the
introduction of the Core 2 Duo (from 95 to 180 million dollars monthly). 49% of that came
from the introduction of new products (holding old products fixed), and the rest came from
the realignment of products throughout the spectrum. Finally, a monopolist with the same
innovation would have increased profits by 17% (from 488 to 573 million). Even though a
monopolist has lower percentage returns, in dollar values, the amount is very similar to the
oligopoly outcome.
These results speak to recent antitrust enforcement in this industry. The market leader,
Intel, has been widely accused of actively working to exclude AMD from the market. A

the second-stage pricing game and estimation of marginal and sunk costs. Section 5 lays
out and solves counterfactuals using the estimates from earlier sections. The counterfactuals
simulate what the market would look like if it were a monopoly, run by a social planner, or
had different innovation outcomes.
5
Eizenberg shows how to account for issues of self-selection and partial identification in these sorts of
games, an estimation problem closely related to the one in this paper.
6
2 The CPU Market
The market for desktop, laptop, and server CPUs is dominated by two companies: Intel and
AMD, with (respectively) approximately 80% and 18% market share as of January 2009.
This paper concentrates on the market for desktop CPUs. These are CPUs that go into
home and business machines that are used for everyday tasks. I concentrate on this market
rather than the market for laptop chips because it is more competitive (Intel dominates the
market for laptop chips) and more stable (laptop growth has been explosive over the last few
years). More data are also available for this market because enthusiasts tend to buy desktop
chips and chart their performance extensively.
Within the desktop market, each firm typically offers between 10 and 15 chip varieties
at any given time. By far the largest difference between these chips is performance. Higher
performing chips tend to have higher clockspeed (operate at a higher frequency), more high-
speed cache memory available, include multiple cores, and use more advanced process tech-
nology. Firms can and do use all of these levers to manipulate performance, but at the end
of the day a consumer need only look at how the chip performs based on some benchmark
to determine it’s product quality.
6
The CPU market has long been known for offering quality-differentiated product lines.
The Intel 80486 was a popular example in both the economics literature (Deneckere and
McAfee (1996) ) and the popular press. Introduced in 1989, Intel created low-quality and
high-quality versions of this chip. Strikingly, in order to create the low-quality chip, they
went to some cost to destroy a perfectly good high-quality chip.

The competitive nature of the industry has fluctuated markedly over the 2000’s. The
mid 2000’s were the height of AMDs competitiveness, peaking with 30% marketshare at the
beginning of 2006. This is up from around 10% at the beginning of 2003 and 20% at the
beginning of 2009. Figure 1 plots Intel and AMD marketshare from the end of 2003 through
the beginning of 2009.
[Figure 1 about here.]
Changing technical leadership is partially responsible for the marketshare fluctuations. From
2002-2006, AMD consistently released products whose price/performance characteristics
were similar to or beat Intel’s products. However, with the release of the Core 2 prod-
uct line at the end of 2006, Intel regained technical leadership, a position which they’ve held
every since.
Figures 2 and 3 tell the story of a rapidly and significantly changing market. In June
2006, Intel and AMD both offered products throughout the quality spectrum. In many
parts of the spectrum, AMD offered products that were better and cheaper than comparable
Intel products (the top panel of 2). In July 2006, Intel released a number of Core 2 Duo
8
See: I thank Kelly Shue for pointing this out to me.
8
products. The bottom panel of figure 2 shows that these products completely dominated
AMD’s offerings, substantially altering the price/quality landscape. AMD responded by
slashing prices and removing products that were no longer competitive (figure 3). By January
of 2008, the competitive nature of the market had changed so significantly that AMD was
relegated to the bottom portion of the quality spectrum, offering almost no chips at the
medium to high end. Interestingly, while AMD’s overall marketshare did not drop all the
much. Its share of more expensive chips dropped almost to 0 (figure 4).
[Figure 2 about here.]
[Figure 3 about here.]
[Figure 4 about here.]
I exploit this shift in technological leadership in two ways: First, it provides natural
variation in product characteristics that helps identify the demand system. Second, I explore

tion of CPU speed and cache) because I am explicitly comparing performance across CPU
manufacturers with substantially different architectures. Simple clockspeed doesn’t reflect
actual performance in this case because the chip architecture interacts with numerous char-
acteristics of the chip in complex ways to actually move information through the pipeline.
9
There is plenty of evidence that the large OEMs do not pay these list prices. Instead, at a minimum,
they get percentage discounts off of them depending on their size. As long as these are the same across
the OEMs in my data, then my substantive empirical conclusions will not be affected. It would be more
problematic if specific OEMs got specific discounts off of certain chips and not on others. I have not heard
of instances of this occurring, but given the bilateral nature of these deals, I certainly cannot rule it out.
10
3 Demand
Demand follows the Pure Characteristics Model of Berry and Pakes (2007) (See also Song
(2007)).
10
Consumers choose a single product from the set of available options. Conditional
on purchasing, consumers get utility as specified in equation 1.
u
ij
= β
i
x
j
−
1
α
i
p
j
+ ξ

+ ν
i
d
j
−
1
α
i
p
j
+ ξ
j
(2)
I allow for consumers to have varying tastes for purchasing an Intel or AMD product by
including the random coefficient, ν
i
on d
j
. If an Intel product, ν
i
will enter the utility
function positively and if an AMD product, negatively. ν
i
determines how substitutable
products are across firms. At one extreme, if ν
i
is a constant with value 0, then two products
(from different firms) with the same characteristics would be identical from a consumer’s
10
The key difference between this model and the more commonly used discrete choice model of Berry,

buying decision. Second, due to differences in architectures, AMD and Intel chips perform
different types of tasks at different speeds. For instance, those who play computer games
tend to prefer AMD chips, while those with business oriented tasks prefer Intel. Lastly, there
are complementarities between the CPU and other components inside the computer. CPU’s
from a given company can often be upgraded without changing the motherboard, video card,
etc, while changing to a CPU from a different company would require re-purchasing those
components.
Splitting apart the random coefficients into a mean and variance term and assuming that
α
i
is distributed lognormal and ν
i
normal gives (where n
ν
and n
α
are standard normals):
u
ij
= βx
j
+ (¯ν + σ
ν
n
ν
)d
j
−
1
exp(¯α + σ

u
ij
= δ
j
+ σ
ν
n
ν
d
j
−
1
exp(σ
α
n
α
)
p
j
(5)
3.1 Estimation
The parameters to be estimated are the mean utility levels (δ
j
), and the variances of the
random coefficients (σ
α
, σ
ν
). α
i

k
− δ
j
(6)
Then all consumers with:
∆
j
n
< α
i
< ∆
j
n
(7)
Will purchase product j.
Here I diverge from the original Berry and Pakes estimation routine. Because of the
structure of my model, one horizontal and one vertical characteristic with two firms, I am
able to calculate an analytical cutoff value for each α
i
type that determines which consumers
purchase from each firm. Let u
j
∗
n
be a consumer’s preferred product conditional on purchasing
from firm n. Then, for every α
i
, there is a cutoff value, ˜ν, such that u
j
∗

1(2) are given by:
S
j
1
=
∆
j
1

∆
j
1
[1 − G (˜ν(p, δ, α)|α)] dF (α) (9)
S
j
2
=
∆
j
2

∆
j
2
[G (˜ν(p, δ, α)|α)] dF (α) (10)
This formulation is helpful for two reasons: First, because I don’t need to simulate consumer
types in order to calculate predicted marketshares (I can use numerical quadrature), I re-
duce noise in the model. Second, the relative ease of computing marketshares allows me
to more easily formulate the problem as a constrained maximization problem and use the
Mathematical Programming with Equilibrium Constraints (MPEC) techniques as discussed

ν
,σ
α
,
˜
δ,p,β
˜
ξ(β, σ)

ZW Z

˜
ξ(β, σ) (11)
Subject To:
s
j
= S
j
∀j (12)
For each candidate value of the standard deviations of the random coefficients, the routine
calculates the mean utility levels that equate actual with observed marketshares and finds
the value of the objective function by taking the residuals from an IV regression. The
IV regression has the mean product utilities on the left-hand side and the right-hand side
includes product dummies (k
j
), time dummies (k
t
), and market dummies (k
m
). See equation

jtm
(13)
This estimation routine presents a computational challenge: Unlike in BLP, where the id-
iosyncratic error term ensures positive market shares for all products at all potential param-
eter values, this model often predicts that some products will have 0 market share. This
11
The downside to this estimation strategy is that recovering the underlying utility coefficients requires
an extra step of regressing the dummies on product characteristics – a regression that is only valid if the
unobserved product characteristics are uncorrelated with observed product characteristics. In this context
this means firms don’t know ξ when they choose product characteristics, an assumption that is unlikely to
be true in my case. Fortunately, I’m not actually concerned with the coefficients on product characteristics
– recovering the δ

s is sufficient for the supply side estimation and counterfactuals.
15
occurs when the mean utility levels become “un-ordered,” making some products dominated
by their neighbors. When this happens, the gradient of the constraints for those products
are 0 and standard computational techniques no longer work. However, for parameter val-
ues where the marketshares are all non-0, this is a straightforward constrained maximization
problem. The key, then, is to get good starting values. To do this, I implement a routine that
smoothes out the marketshare constraints by viewing consumer’s choices probabilistically.
12
I then use these as starting values for the analytical market share routine. I solve this as
a mathematical program with equilibrium constraints (Su and Judd 2008) using the opti-
mization routine KNITRO where marketshares are computed with quadrature. Appendix 2
details this process.
3.2 Instruments
Instruments other than the X’s themselves are necessary both because I estimate the standard
deviations of the random coefficients (requiring at least two extra instruments) and because
firms choose prices knowing the demand shocks, leading to potential correlation between

this informal ranking.
15
The right-hand panel of 1 shows a selection of month dummies, the
country dummies, and the estimates of the standard deviations of the random coefficients.
This industry is highly seasonal resulting in month dummies that are higher in December
compared to the rest of the year.
[Table 1 about here.]
4 Quality Choices
Supply side competition is assumed to proceed in two stages. In a first stage, firms choose
product qualities, paying a sunk cost for moving products from their previous spot (products
are moved by introducing a new product and taking out an old one that was at the same
price level). In a second stage, firms compete in prices taking quality choices as given.
13
I list these as being illustrative examples that come out of the estimation. A table listing the quality
levels of the 91 chips that exist in my data would not add all that much and would significantly detract from
the readability of the table.
14
That is to say, by chip generation in the order of better chip generations and by product number within
chip generation because product numbers often give a rough idea of how the CPU companies themselves
view which products are better than others.
15
It is somewhat interesting to see that the high end products from lower chip lines sometimes are valued
higher than the low end chips from higher chip lines (compare, for example, the Celeron 360 and the Pentium
4 531).
17
In reality, product quality choices are dynamic. Because products live for more than
one period, their placement today affects sunk costs that would need to be paid tomorrow
should they be moved. I follow most of the literature in assuming a two-period model for a
number of reasons: First, the main point of this paper concerns ways in which competition
interacts with product placement decisions, and to understand these incentives, dynamics

1
MS
mj
1
(p, δ) − C
j
1
(14)
Country: m, Product: j
The maximization problem assumes that prices don’t vary from country to country (arbitrage
would prevent prices from diverging too much) and that marginal costs are the same across
countries.
16
16
The p’s and C’s have product but not country subscripts and the problem for the firm maximizes over
18
This yields a first order condition of:
∂Π
1
∂p
j
1
=
M

m=1
(p
(j−1)
1
− C


j
1
)
∂S
mj
1
(p, δ)
∂p
j
1
+ S
mj
1
(p, δ)
(15)
The equilibrium is a fixed point in p for the two firms.
The own price derivative for product j is given by:
∂S
j
1
∂p
j
1
=

1 − G(˜ν(p, δ, ∆
j
1
)|∆

1
∂p
j
1
−
∆
j
2

∆
j
2
g(˜ν(p, δ, α)|α)
∂˜ν(p, δ, α)
∂p
j
1
f(α) dα
(16)
This is a very natural equation. The first two terms quantify the consumers that are lost
to the products above and below in the product space. However, not all consumers are lost:
only those that were already purchasing products from firm 1. If all consumers purchased
from firm 1 so that G(˜ν(p, δ, ∆
j
1
)|∆
j
1
) = 0, then the model collapses back to the vertical
model. The third term quantifies consumers that are lost to firm 2 through a change in ˜ν.

1
(17)
the sum of profits from the individual countries (m).
19
4.1.1 Marginal Costs
Marginal costs are assumed to follow a Markov process. A chip’s cost in any given period is
a function of last period’s cost and an idiosyncratic error term:
c
j(t+1)
= β
j
+ ρc
jt
+ 
c
(jt)
(18)
Given the demand parameters, the first order conditions (equation 15) define a set of non-
linear equations that can be used to back out the marginal cost of a chip in any period.
17
Estimation begins by solving these systems of equations period by period, giving estimates
of each chip in each period.
18
These estimated costs are then regressed on last period’s costs
to get an estimate of ρ.
Figure 6 graphs estimated marginal costs for AMD and Intel for two months: May
2006 and May 2007. At the low end of the product spectrum, costs for AMD and Intel
are relatively similar, but AMD’s costs rise faster with quality than Intel’s, resulting in
significantly higher costs for producing high quality chips. This asymmetry in costs is a
large part of Intel’s competitive advantage. Between May 2006 and May 2007, costs for both

a broad closely-spaced product line. The counter-balancing of these forces will determine
the equilibrium.
The formal problem for firm 1 is laid out in equations 19 and 20.
max
δ
1
E

π(δ
1t
, δ
2t
, p
∗
1t
, p
∗
2t
,
˜
ξ
t
, 
ct
)

− SC(δ
1(t−1,t)
) (19)
max(δ

introduce a new product. If they introduce it, then it must have been the case that it was
more profitable than not introducing it and not paying the sunk cost. Similarly, if they
decide not to introduce a product, then it must have been the case that the firm would
have been worse off introducing the product and paying a sunk cost. These two optimality
conditions allow me to implement an inequality estimator in the style of Pakes, Porter, Ho,
and Ishii (2006).
Firms are assumed to pay a constant sunk cost for introducing a new product irrespective
of where in the quality space that product is located. Letting δ
jp
denoted the location of
products in the previous period, sunk costs are given by:
SC =
J

j=1
I(δ
j
= δ
jp
)θ (21)
This procedure uses the estimated demand system to calculate a pricing equilibrium and
consequent profit for actions that the firms could have taken but decided not to take. On
average, observed actions are assumed to be more profitable than the unobserved potential
actions. The difference between profit from observed and unobserved actions is used to
form moments. Without further assumptions, sunk costs are not point identified, rather the
moment conditions allows me to identify bounds.
One side of the bound comes from looking at products that a firm could have moved (by
adding a new product to the line and removing the old product) but decided not to. If they
moved the product, they would in expectation increase profits but also incur the sunk cost
of product introduction, θ. Letting π(δ

tations and realized profit. I assume that ν
sc
is unobserved by both the econometrician and
the firm.
20
Denoting r(δ
j
) as the estimate of observed profit that comes from the demand
and costs estimates, the relationship between π(δ
j
) and r(δ
j
) is given by:
E [π(δ
j
)] = r(δ
j
) + ν
sc
(24)
As long as firms have correct expectations on average, ν
sc
will go to 0 as the sample size
goes to infinity. This gives:
plim
J→∞
1
J
δ
j

, the strategy that a firm could have used but decided not to, I find the best possible
placement for a product by searching across the δ space and recomputing the competitive
pricing equilibrium for each possible spot.
Using this procedure, I estimate that the sunk costs of introducing a product fall in the
range of $1, 236, 000 − $3, 412, 000. This is rather small compared to profits but consistent
with the idea that once a chip generation is introduced, adding an additional product doesn’t
require a whole lot of extra work.
5 Counterfactuals
Putting together the estimates described above with the structure of the model, allows me
to construct counterfactuals that speak to the role that competition plays in this market.
20
Pakes, Porter, Ho, and Ishii allow for a second error that firms know but that is unobserved by the
econometrician. With appropriate instruments, this can be included in the model. I don’t allow for this
second error, implicitly assuming that sunk costs are the same across firms and across time.
23
Key parts of the counterfactuals examine how consumer welfare changes under different
scenarios. To be specific about this, consumer welfare is given by:
CS
intel
=
∆
j
1

∆
j
1
α
i
ν

∆
j
1
α
i
˜ν

ν

δ
j
−
1
α
i
p
j
+ ν
i

f(ν
i
)f(α
i
) dν
i
dα
i
(27)
Where CS

i
distribution.
24
introduced, monopolists will choose different quality levels.
22
The goal of this counterfactual
is to separate out the different components of consumer welfare change and quantify potential
social loss or gain.
The first set of counterfactuals, detailed in table 2, decomposes a shift to monopoly into
the profit gains and welfare losses from each of these mechanisms. Removing AMD but
fixing products and prices doesn’t change profit (up 2.8%) or consumer welfare (down 1.2%)
very much. The relatively small change in consumer welfare indicates that consumer taste
for AMD products is not all that strong. Individuals who were purchasing AMD products
suffer some loss from purchasing an Intel product or substituting to the outside option, but
their lost utility is not large.
[Table 2 about here.]
Next, I solve the monopolist’s profit maximization problem, keeping product qualities fixed
at the competitive level. Figure 8 plots markups as a function of quality (δ). Prices rise
and Intel’s profit goes to $562.2 million. Meanwhile consumer welfare drops 51% to $791.3
million. The increase in prices is telling: It indicates that despite AMD’s relatively small
marketshare, this industry is quite competitive. Removing AMD from the market would
allow Intel to significantly raise prices leading to large consumer welfare losses.
[Figure 8 about here.]
5.1.1 Optimal Product Placement
Consistent with equation 19, a monopolist solves for optimal quality choices knowing the
markov process that costs follow (and the ρ parameter of that process) but without knowing
22
In theory not all consumers are necessarily made worse off by a monopolist: if quality levels change such
that consumer types are served that weren’t served under oligopoly or prices go down on some products in
response to monopolist segmentation, then welfare for those consumers could go up. In practice, it is highly