Operating Efficiency in the Capital-Intensive Semiconductor Industry: A Nonparametric Frontier Approach

1 Introduction

Integrated circuits (ICs) are essential components of virtually all modern electronic devices. Since Bell laboratories invented the transistors in 1947 and Texas Instruments released the first working IC in 1958, the semiconductor industry, which is the aggregate of companies engaged in the design and fabrication of semiconductor devices or IC chips, has been at the forefront of the digital economy for decades. From laptop to smartphone and artificial intelligence (AI), semiconductor devices are present in nearly all aspects of modern technology. The personal computer revolution in the 1970–1980s was a result of advances in semiconductor technology, such as the Intel 8,008 microprocessor (Ceruzzi, 1996). In the development and expansion of the World Wide Web revolution in the 1990s, application-specific integrated circuits played a significant role in enabling fast and efficient networking and data processing, contributing to the growth of web-based technologies and consumer electronics (Makimoto, 2002). The rise of the smartphone in the 2010s was supported by higher-performance system-on-a-chip (SoC), such as the Apple A series and Qualcomm Snapdragon series. ChatGPT, latterly the most popular deep learning workload, required significant computational power and was trained on Nvidia^[1] graphics processing units (GPUs).

Due to the growth in emerging technologies such as AI, cloud computing, Internet of Things, 5G networks, autonomous vehicles, industrial automation, and renewable energy systems, the needs for more powerful, energy-efficient, and miniaturized semiconductor devices have been consistently increasing. The 2023 Semiconductor Industry Association (SIA) factbook reported that the global semiconductor sales reached the highest-ever annual total of $574 million in 2022.^[2] McKinsey’s preliminary forecast shows that the global semiconductor industry is poised to become a trillion-dollar industry by 2030 (Burkacky et al., 2022).

2 Methodology

3 Data and Variable Specification

The data are collected from the Sub-Industry of Semiconductors in the Compustat database. In order to provide a global perspective for the semiconductor value chain, we combined data from both the Compustat North America database and the Compustat Global database to cover companies in the industry worldwide. As the semiconductor industry is famous for being a cyclical industry (e.g., see Rastogi et al., 2011; Tan & Mathews, 2010), we gather 20 years of data in 1999–2018 to cover a sufficient period with multiple business cycles in the industry. We exclude liquid crystal display manufacturers, light-emitting diode manufacturers, and photovoltaic producers from the dataset, limiting the sample to only IC manufacturers in a narrow sense. Hence, the panel data include 5,136 observations from 470 unique companies in the global semiconductor industry in 1999–2018.

The reason for the data to begin in 1999 is twofold. First, the global semiconductor value chain has been preliminarily established in the late 1990s since the inception of the fabless–foundry business model in the late 1980s. Since 1999, there are plenty of available annual reports for the fabless and foundry firms on the open market. Second, the year 1999 is a suitable starting point to observe the development trend in the global semiconductor industry. Two years after the 1997 Asian financial crisis, the semiconductor industry is in a golden decade without massive exogenous shocks until the 2008 financial crisis.

Identifying the inputs and outputs of the production function has always been a subject of controversy, either in parametric or in nonparametric frontier estimations, without exception in the semiconductor industry. Hence, we sort the most commonly used variables in 37 empirical studies, which apply the DEA approach in the semiconductor industry. Besides a few variables chosen for specific topics, the commonly used variables in these articles are highly concentrated into two input and two output categories. The first input category measures variable inputs, including labor, raw material, R&D, sales, and marketing expenditure, while the second input category measures fixed assets. Therefore, we specify p = 5 inputs (labor, measured by the number of employees [ X 1 ]; COGS [ X 2 ]; R&D expenditure [ X 3 ]; sales and marketing expenditure [ X 4 ]; and fixed assets, measured by property, plant, and equipment [PP&E] [ X f ]). Note that we distinguish the notation of the fixed input X f from the other variable inputs X 1 , X 2 , X 3 , and X 4 .

Comparably, the first output category measures revenue and the second output category measures the market value of the firms. Hence, we specify q = 2 outputs (total revenue [ Y 1 ] ; and shareholders’ equity, measured by common ordinary equity [CEQ] [ Y 2 ] ). For the output variable Y 2 , we use shareholders’ equity instead of the market value of a firm, because the variable of market value is suffering from missing data in the Compustat database, and the variable shareholders’ equity is also a widely used proxy for the value of a firm. The same dataset had been used in Qiao and Wang (2021), but in this approach, we add the variable X f to emphasize the impact of CAPEX, and measure it as a fixed input by the directional distance estimator to achieve more robust results.

Table 1 gives the summary statistics for the variables in 1999–2018 pooled data. In order to provide a uniform standard across years, all the variables except X 1 are expressed in US $ millions, and their values have been adjusted to 2018 US dollar by GDP deflator. The distribution of all the variables is heavily skewed to the right, owning to the domination of several semiconductor giants in the market.

Besides inputs and outputs, we specify r = 2 environmental variables (business model [ Z 1 ]; and time, measured by the years 1999–2018 [ Z 2 ]). The business model Z 1 is a discrete variable, which categorizes the business models of fabless, IDM, foundry, and A&T into three groups. The first group contains the fabless, which are labor-intensive for IC design. The second group contains foundries and OSATs, that are capital-intensive for fabrication. The third group contains IDMs that are both labor-intensive and capital-intensive. Alternatively, Z 2 can either be a continuous variable or a discrete one. If Z 2 is treated as a continuous variable, choosing the optimal time windown for Z 2 is critical, which will be discussed further in the next section.

Table 2 breaks down the 5,136 observations by the business model. It is no surprise that over half of the companies are fabless. As the barriers to entry, which rely heavily on CAPEX, are much lower for fabless than for the others, fabless companies spring up like the mushrooms in the late 1990s to the early 2000s. At the same time, the number of firms operating in other kinds of business models remains relatively stable. After the golden decade of fast growth in the semiconductor industry come to an end in the mid-2000s (e.g., see Flamm, 2017), the proportions of firms in each business model are gradually fixed. Around 60% of the firms are fabless, while 20% of the firms are IDMs and the rest 20% are either front-end wafer fabs or back-end OSATs.

4 Empirical Results

Most nonparametric estimators suffer from the curse of dimensionality. Based on the three diagnostics introduced in Appendix A, the necessity for dimension reduction is unambiguous. With seven dimensions ( p = 5 and q = 2 ) in the original data, the effective parametric sample size m for annual data is small, no matter using FDH or DEA estimators. We calculate the values of the largest eigenvalue of the moment matrices of X X ′ and Y Y ′ to the corresponding sum of eigenvalues to be R x = 91.19 % and R y = 98.31 % , indicating high correlations among the inputs and among the outputs. Therefore, dimension reduction can reduce estimation error. A slight difference in processing PCA for the directional distance estimator is that PCA is only on the variable inputs and outputs, but not on the fixed input X f . Hence, after PCA, there remain three dimensions, namely, X ˜ (PCA from X 1 to X 4 ), X f , and Y ˜ (PCA from Y 1 to Y 2 ). The following analyses are based on data with PCA.

Among studies that use the nonparametric frontier approach to estimate efficiency and benchmark performance of firms in the semiconductor industry, the vast majority choose the DEA estimator, without comparing the pros and cons between the FDH estimator and the DEA estimator. The DEA estimator is probably a better choice without dimension reduction, as the slower convergence rate of FDH estimator may increase measurement error rapidly with increasing dimensions. However, it is worth to reevaluate the trade-off between the FDH and DEA estimators with dimension reduction. The drawback of the DEA estimator is imposing convexity on the production set Ψ , while the FDH estimator is free of this assumption. The test of convexity is applied to measure this trade-off.

Table 3 provides the results of the convexity test.^[21] At 95% confidence level, the null hypotheses of convexity are rejected for over 80% of the 20 years’ annual data, except 3 years (2009, 2011, and 2012) in the hyperbolic measure. Simar and Vanhems (2012) linked the directional distance measure with the hyperbolic measure by a monotonic transformation, so that the results in Table 3 are also valid for the directional distance estimator. Hence, we choose the FDH estimator.

In order to consider a discrete environmental variable such as the business model Z 1 for the estimator in equation (2.4), the separability condition needs to be examined. We use a bootstrap algorithm (Simar & Wilson, 2020) for the separability test on the discrete environmental variable Z 1 .^[22] The first portion in Table 4 show the separability test results with respect to the business model Z 1 . Though the test statistics τ 1 and τ 2 not always give the same results, there is strong evidence to reject the separability condition. In other words, each of the three business models in semiconductor industry has a unique production frontier.

For the environmental variable Z 2 , which represents the years 1999–2018, there is flexibility to treat it either as a discrete variable or as a continuous variable (Mastromarco & Simar, 2015). To treat Z 2 as a discrete variable, the 20 years of 1999–2018 can be splitted into ten 2-year groups (two adjacent years as a group), five 4-year groups (four adjacent years as a group), or four 5-year groups (five adjacent years as a group). In this case, the separability test with respect to Z 2 is similar to the separability test with respect to Z 1 .

Although Z 2 can be treated as 20 individual years, it is not recommended. From the economic point of view, the technology life cycle and business cycle in the semiconductor industry are typically 2–4 years (Tan & Mathews, 2010), so it is more natural to assume a uniform production frontier for the semiconductor companies within one industry cycle. From the econometric point of view, since there are 100–300 observations in each year, the effective parametric sample size m = n 2 3 for the directional distance measure in each year is very small. In either way, the approach to treat Z 2 as 20 individual years is less attractive.

Alternatively, treating Z 2 as a continuous variable, the optimal time window of Z 2 needs to be fixed in advance. The optimal bandwidth is h = 5.5 for fabless, IDM, and pooled data, implying that the smoothing window of year t is [ t − 5 , t + 5 ], while the optimal bandwidth is h = 7.5 for OSAT with the smoothing window of year t to be [ t − 7 , t + 7 ] (Appendix C explains how to calculate the optimal bandwidth). The second portion and the third portion in Table 4 show the separability test results with respect to the optimal time Z 2 , and with respect to the business model Z 1 and the time Z 2 , respectively. In any case, the separability conditions are strongly rejected. Hence, we will estimate the efficiency scores defined in equation (2.4) with separated production frontiers per the constraints of both Z 1 and Z 2 .

Table 5 shows the summary of the efficiency scores conditional on both the business model Z 1 and time Z 2 . Whether the time Z 2 is treated as a discrete variable or a continuous variable, the distributions of the efficiency scores are skewed to the right in all kinds of business models, especially for the fabless firms, implying more intensive competition for the fabless. Nevertheless, on conditions that Z 2 is treated as a discrete variable, the first quartiles are either equal to zero or very close to zero, no matter how the years are grouped. It is a sign that the measurement error by slow convergence rate still exists (e.g., see the second diagnostic in Appendix A). Thus, to treat Z 2 as a continuous variable is preferred, rewarding a larger subsample size and a faster convergence rate, and hence more accurate estimates.

Figure 1 visualizes the trends of the annual mean efficiencies by the business model and year. Treating Z 2 either as a discrete variable or as a continuous variable, the curves of the annual mean efficiencies for the fabless firms are above the curves for the other business models. This phenomenon is more visible in the bottom-right panel, where Z 2 is defined as a continuous variable. As a higher efficiency score infers a lower technical efficiency in the directional distance measure, the curves in Figure 1 imply that the fabless firms are operating less efficiently on average.

Figure 1

Mean β conditional on business model and year.

Figure 1 also demonstrates that if Z 2 is a discrete variable, the differences of operating efficiency among business models are fading away over time. Driving by increasing complexity of ICs, technological convergence, and time-to-market and cost optimization pressure, the fab-lite model became popular in the semiconductor industry in late 1990s and early 2000s, and has continued to evolve and gain prominence. Some IDMs and fabless companies are forming strategic partnerships and collaborations to leverage each other’s strengths. Furthermore, associations such as SEMI (Semiconductor Equipment and Materials International), SIA and ESIA (European Semiconductor Industry Association), contribute to the establishment of industry standards and provide platforms for networking and collaboration among industry stakeholders. The boundary between IDMs and fabless companies is becoming increasingly ambiguous.

M&A activities in the semiconductor industry also contribute to the blurring of boundaries between IDMs and fabless companies. The industry has seen numerous mergers, acquisitions, and strategic partnerships aimed at expanding product portfolios, accessing new markets, and driving innovation. On the one hand, fabless giant Qualcomm has made over 41 acquisitions and 108 investments, such as the acquisition of Atheros (a leading provider of wireless networking solutions) in 2011 and CSR (a England-based fabless known for its wireless chips) in 2015, continuously diversifying its product portfolio and strengthening its position in the mobile and networking markets. The merger of Avago and Broadcom in 2015 combined Avago’s expertise in analog and mixed-signal semiconductor solutions with Broadcom’s strengths in connectivity and networking technologies. On the other hand, IDM giant Intel has so far acquired more than 90 companies, including the acquisition of Altera Corporation (a prominent manufacturer of programmable logic devices) in 2015. This acquisition enabled Intel to integrate Altera’s FPGA technology with its processors, offering customized solutions for various applications.

Another interesting discovery is that the curves of different business models in the bottom-right panel of Figure 1 tend to converge in 2008, the year of global finanical crisis. That is, under a severe condition such as the 2008 financial crisis, the differences in operating efficiency become unconspicuous among business models. However, it is important to note that different business models still play a role in determining how companies weather the crisis. While the differences in operating efficiency might not be as apparent during extreme conditions, they can still influence a company’s ability to adapt, survive, and eventually recover. As the economy recovered from the 2008 crisis, the differences in operating efficiency among business models once again became noticeable in the semiconductor industry.

Figure 2 shows the annual mean efficiency curves by business model and optimal time, with 95% confidence interval. The confidence interval is derived by new central limit theorem (Kneip et al., 2015, p. 409). The variances for the fabless firms are higher compared with the IDMs or OSATs, implying higher risks and uncertainties for the fabless business model.

Figure 2

Mean β with 95% confidence interval.

As the separability condition does not hold (Table 4), we use a flexible nonparametric location-scale model for a second-stage regression. The pure efficiency can be derived from equation (2.5). In practice, we obtain μ ^ ( z ) by regressing β ^ ( x , y ∣ z ) on the environmental variable z . Similarly, we obtain σ ^ ( z ) by regressing the squared residuals of the preceding regression on z . The upper panel in Figure 3 illustrates the pure efficiency ε ^ ( z 1 , z 2 ) that cleanses efficiency scores from the influence of both the environmental factors Z 1 and Z 2 . In the upper panel, the curves of ε ^ ( z 1 , z 2 ) by different business model twist together with no clear structures, similar to white noise vibrating at small values around zero.

Figure 3

Pure efficiency.

The lower panel in Figure 3 illustrates the pure efficiency ε ^ ( z 2 ) that cleanses efficiency scores from the influence of only Z 2 , but not the business model Z 1 . In the lower panel, the curves of ε ^ ( z 2 ) demonstrate clear separation by business models. Since ε ^ ( z 2 ) only cleanses the influence of time, the lower panel in Figure 3 maintains the structure of the differences in technical efficiency by the business model in Figures 1 and 2. The contrast between the upper and lower panels in Figure 3 provides further evidence that the technical efficiencies do vary in the semiconductor by the business model, and the asset-light fabless firms are operating less efficiently on average in the past two decades.

5 Summary

The semiconductor industry is famous for the high barriers to entry, especially in the capital-intensive manufacturing portion. The incumbent IDMs, benefitted by the economy of scale and protected by the economic moat by huge CAPEX, have dominated the semiconductor industry since the onset of the industry in the 1960s. However, wagering on novel technologies and processes with the ever-expanding complexity of ICs becomes a weighty burden even for the giant IDMs. The fabless–foundry business model alleviates the financial risks of capital investment, reduces the barriers to entry, accelerats technology iterations, and leads to a flourishing of fabless design houses for various applications.

This study compares the operating efficiencies between the IDMs and the fabless–foundry business models to shed light on which business model will be the market trend and dominate the semiconductor industry in the long run. Based on the capital-intensive feature of the semiconductor industry, this study chooses a directional distance measure to handle the constraint of CAPEX. At the same time, conditional FDH estimators are used to handle the effects of business model and time in the nonparametric frontier approach. The empirical results provide clear evidence that the IDMs are operating more efficiently, while fabless firms are operating less efficiently by and large. A second-stage nonparametric location-scale regression is used to further check the robustness of the finding. By setting different conditions in the second-stage regression, we show strong evidence that the technical efficiencies do vary in the semiconductor by the business model, and the asset-light fabless firms are operating less efficiently comparing with the capital-intensive IDMs, foundries, and OSATs on average.

Nevertheless, there are several limitations in this study. The method is mainly econometric. Criteria for PCA, the separability condition, the optimal bandwidth, etc. are statistical, lacking solid ground in economics. The dynamics in the semiconductor industry had not been fully measured, which may need more advanced econometric methods and data with more technical details for further research.

Though the fabless–foundry business model encourages entrance of the fabless startups, the CAPEX barriers accompanying with technical barriers still limit the fields and applications for the fabless firms to growth and development. The IDMs, having more room to optimize the operation and lead the technology development with a strategic product roadmap by vertical integration, will continuously dominate the semiconductor industry in the foreseeable future. At the same time, the fabless–foundry business model is an important complement of the IDMs to explore a broader scope in the semiconductor industry. In fostering a healthy and competitive semiconductor ecosystem, we suggest to tilt the subsidies of the semiconductor industry toward the fabless sector to promote innovation, to support entrepreneurship and startups, to leverage specialization and expertise, and to encourage more diversification.