Productivity and Cyclicality in Semiconductors: Trends, Implications, and Questions: Report of a Symposium (2004)

David C. Mowery

University of California at Berkeley

Dr. Mowery said that one of the motivations for the agenda of the meeting was to try for a better understanding of cyclical behavior, especially in the semiconductor industry, which was then in a severe downturn. He said that he had asked the next three speakers to address at least four issues:

1. the kinds of data or evidence needed to address causes and characteristics of cycles in these industries over time, including changes in the cycles’ amplitude and duration, as well as data that allow one to develop time series or longitudinal analyses of these cycles;

2. the linkages, if any, between cyclical behavior in a given industry and the behavior of other forms of investment, particularly in physical capital or innovation R&D;

3. the managerial strategies developed in these industries to deal with these cyclical fluctuations, and how these strategies are influencing the evolution of industry structure and investment in areas such as physical capital and R&D; and

4. keeping in mind that the conference was organized by the STEP Board, the implications for public policy of the analysis of cyclical behavior in a given industry, and whether there is a role for government in dealing with the causes or consequences of industry cycles.

Kenneth Flamm

University of Texas at Austin

Dr. Flamm said he would address the “stylized facts” that should be considered in trying to model the semiconductor industry, as well as some of the modeling work under way at SEMATECH and the University of California at Berkeley. He began by listing some key economic features of the semiconductor industry that might be associated with deep cyclical swings.

The first is rapid technical progress. This means that holding inventories as a way of smoothing out demand fluctuations is not productive; because of the short lifetimes of semiconductor products, the inventories will lose value and even become obsolete if they are held too long. Instead, firms must sell what they have as quickly as they can.

Second, very large R&D investments are required to enter this industry—typically, as much as 10 to 15 percent of annual sales. This R&D is often specific to the segment of the market that the firm is entering.

Learning economies are very important. For military aircraft, a learning curve of 85 to 90 percent is estimated, which means a doubling of output drops unit costs by 10 to 15 percent. In semiconductors this curve is closer to 70 percent, quite a bit steeper. In addition, the source of the learning economy in semiconductors is different from that in aircraft.

The curve is not caused necessarily by labor productivity. Instead, improvements come from two sources. One is die-shrinks: Over the cost of a product cycle, the number of chips on a wafer increases. Over any product’s life, this happens typically two or three times, essentially increasing the product on each silicon wafer. The other source is yield learning: The number of good chips on the wafer increases over time as a percentage of the total number of chips. Together these sources generate the steep learning curve. This curve is thought to have flattened somewhat, but good evidence of this is difficult to come by.

Capital intensity is very high and rising. A mid-sized fab today costs $1.5 billion to $2 billion, or even more. Furthermore, if a firm decided to build a fab and then wanted to exit the industry, the resale value would be low. Also, it typically takes 1 to 2 years to build a new fabrication facility. This long gestation time is often overlooked in addressing cyclicality. A spike in demand might prompt a decision to build a fab, but it would not be ready for mass production for a year and a half or two years. This lag time plays a significant role in the very wide swings in the industry, because demand might be fading just as new capacity reaches the market.

The importance of capacity constraints, too, is often overlooked. The output of the semiconductor industry is really a mixed bag including both old, trailing products and new, leading-edge products. Typically, the leading-edge product is produced in the most modern facilities, which are run at full capacity. A modeling strategy that approaches an optimization problem is likely to run up against a constraint boundary. This is important for anyone trying to use interior first-order conditions to make inferences about, say, marginal costs.

Finally, the sources of deep cyclical swings are difficult to quantify, but conversations with people in the industry indicate that periodic technology shocks have been important in explaining spurts of robust demand. In the early 1990s, for example, the PC market experienced a boom, and this in turn created an unprecedented demand for memory chips that lasted approximately three years. Again, the dawning of the Internet era around 1995 created additional demand for semiconductors that only now appears to have tapered off. Wireless communications also propelled the market forward in a way that had not been anticipated.

Previously, said Dr. Flamm, most published data on capacity utilization of semiconductors were misleading. Typically, published capacity numbers referred to the entire semiconductor industry. The problem with such numbers is that it mixes older and newer products. The older products are typically produced in depreciated, older fabs that run far below capacity.

It has been common to hear that “the industry is running at a capacity utilization of between 40 and 80 percent” and to assume that capacity constraints were not an issue. In fact, between 1997 and 2001 plants manufacturing semiconduc-

tors at feature sizes of 0.4 micron and below, close to the technological frontier, were typically running at 90 to 97 percent of capacity during a period of robust demand.

Another way to view the capacity issue is by analyzing 8-inch wafer starts. This wafer size represented the frontier in 1997, when plants were running at levels above 90 percent of capacity. With the downturn in 2001, the numbers of wafer starts at leading-edge technologies dropped to the 70- to 80-percent range of capacity utilization. But the latest-generation technologies, with features smaller than 0.2 micron, were running as close to capacity as they could.

The reason for this is that the brand-new fabs are the ones that can produce the highest quantities of leading-edge product and do so economically. Even in periods of declining overall capacity utilization, the leading-edge plants are almost always running at essentially full capacity, minus that small fraction of plant equipment that must always be down for upgrade and service purposes.

He then reviewed a project undertaken for SEMATECH to create an economic I/O-style model. To do this, he assumed a vector of demand for semiconductors that was disaggregated into product classes and technology nodes. He worked backward from that vector to detailed requirements for materials, which were driven by an assumption about demand. One goal was to put prices and demand into the model, and another was to create a more realistic investment model, where the amount invested depended on the return on investment (ROI). He also wanted to incorporate some of the influences that produce cyclical behavior, including the general state of the economy. A final goal was to convert the SEMATECH model to an investment algorithm that was based on ROI and could deal with such factors as gestation lag and expectations.

He reviewed a simpler approach to modeling the industry by a deterministic optimal-control model, devised in 1996. That approach assumed that investment in the industry proceeded in two stages. The first was a capacity-investment stage when, at the outset, a firm would invest the amount required to enter the industry, sink resources into capacity, and begin production. This was a relatively simple, open-loop, optimal-control model where every entrant was equal. He found interesting regularities according to the segment of the industry where he applied the model. For DRAMs, the Hirfendahl Hirschman index was strikingly constant from the 1970s through the 1990s at about a level of about 0.1—that is, for a symmetric industry of about 10 equally sized players. That shape seemed to have

changed recently as the industry had become more concentrated. But he completed a prototype version of a Nash equilibrium finder that ran with the Excel solver and hoped eventually to integrate that model with the SEMATECH model, which is also programmed in Excel. So far, he concluded, there has been some success in seeing how changes in some of the parameters affect the determination of equilibrium in the industry.

Asked whether his work had produced any predictions for the industry, Dr. Flamm said that the rates of technological improvement in the semiconductor industry witnessed in the late 1990s did not seem to be sustainable. “I think we may go back to a rate that is somewhat above previous levels,” he said, “but not back to the 60 percent zone we were seeing.”

Dr. Jorgenson asked whether the product cycle can be modeled by Dr. Flamm’s technique—more specifically, whether the model is capable of explaining the product cycle, which appears to drive both the technology and the price, and therefore the demand.

Dr. Flamm replied that he had so far looked only at the small piece of the model representing demand. But he thought that the model might be applied to the business cycle as follows: Suppose a demand function, and suppose that the amount of product demanded will depend on the price per function. That price, coupled with the amount of capacity investment that determines output, would be essentially fixed in the short run, given the output of all firms. This would give some level of return on an incremental investment in output. The return would include not just current price but also an expectation about future prices.

One could think of an iterative algorithm, he said, that would allow calculation of the rate of return on investment. If that rate of return were above the hurdle rate, the firm would want to increase investment. As investment increased, output would increase at every moment in the future, price would tend to decline as all the identical firms in the industry increased their investment, and at some point the rate of return on investment would drop. Investment would essentially cease as it approached the hurdle rate. In this way the model had a mechanism that could take expectations into account.

C. Lanier Benkard

Stanford University

Dr. Benkard talked briefly about cyclicality in the aircraft industry, including a review of some possible similarities with the semiconductor industry.

The first possible similarity he considered was rapid technological progress, to which he said, “Not really.” For aircraft, he said, the basic technology had not changed for many years. One major innovation came in the late 1960s, when high-bypass jet engines were introduced; these essentially made it possible for large modern jets to fly. More recently, the smaller innovation of fly-by-wire technology had not had a large impact to date, although it may enable the next generation of aircraft design.

To a second possible similarity between aircraft and semiconductors, large R&D investments, he said, “Absolutely.” Bringing out a new aircraft product typically requires from $5 billion to $10 billion, which is often more than the market value of the company. Both industries also require ongoing R&D investment.²⁵ Both have learning economies as well. For aircraft the consensus number is 80 percent, he said; the number for Lockheed was about 75 percent.

Both industries have large fixed costs: the cost of buying a plant.

Capacity constraints are very important for aircraft, as for semiconductors, typically increasing during industry booms. These booms do not always coincide with economic expansion, however.

He noted the surprising fact that the largest output year for the commercial aircraft industry was 1991, a year of worldwide recession. This boom was quickly followed by a steep sales slump, then another recovery. Industry sales typically lag the economy, partly because orders must be placed in advance of delivery and partly because labor productivity is low at the beginning of a new product cycle while workers learn their tasks and managers refine production techniques.

One important difference between the semiconductor and aircraft industries is the shape of the learning cycle. The semiconductor production process is capital intensive, and the learning process is primarily a matter of making small, incremental adjustments to the automated technology.

Aircraft, by contrast, are largely “handmade”—that is, they are put together piece by piece. Plants use the labor-intensive practice of training workers to go from station to station and carry out multiple tasks. The reason for this is that the very low unit output rates—plants produce as few as 20 to 40 planes of a given type per year—do not justify the design and purchase of automated machinery for every phase. This means that the learning curve is a function primarily of labor.

One consequence of this labor-based learning cycle in the aircraft industry is that productivity drops late in the cycle. Lockheed, for example, went from a productive level of about 220,000 man-hours per unit to about 450,000 manhours per unit—a 50 percent drop in productivity.

The explanation for this productivity loss for Lockheed is actually one of cyclicality in output. When output goes down, the company has to lay off workers or reassign them to other work stations. Some of the knowledge they have of making aircraft is lost. This “knowledge forgetting” may occur when a manager walks out and doesn’t write down what he knows, or when information that was written down is lost.

He showed a graph that demonstrated the cyclicality for Lockheed. In 1971 the company produced four aircraft. The oil shock of 1973-1974 dragged demand for aircraft down as the price of fuel rose, to claim 50 to 60 percent of the cost of operating aircraft (that portion is now about 25 percent). Demand crashed, and as it started to pick up again, another the oil shock struck in 1979, taking demand down again. In the background of this cyclicality, the productivity of man-hours per unit was highest when output was low, because output had recently been high and workers had reached high proficiency. During low output, they lost proficiency—which had to be regained when sales picked up. This is a countercycle of learning and then unlearning.

Dr. Benkard said he could model Lockheed’s cyclicality with a traditional model of the learning curve. In that model, experience was the equivalent of cumulative past production—the last period’s experience plus the last period’s quantity. He allowed experience to depreciate by a factor delta. When he calculated for delta, he found it to be quite high: about 4 percent per month, or 40 percent per year.

He found several explanations for this high cyclical forgetting. One was that orders are made in advance. Another had to do with capacity constraints. The company would receive orders, ramp up its production lines, and become very efficient through learning. That would have been the most favorable time, when efficiency was high, to maximize production.

Another pattern was shown in the case of Boeing in the 1990s, when the firm’s production fell off in the face of many orders. The company could not increase production because it lacked sufficient experienced workers. The response was to hire even more workers than the company had had before, and this drove its stock price down. The lack of trained workers acted as an implicit capacity restraint.

Dr. Benkard closed by describing how the industry tries to manage these extreme fluctuations by the simple strategy of diversifying into the defense industry. Defense contracts tend to be longer term, and defense production can be accelerated when commercial production is down.

Dr. Mowery asked whether one motive for companies that lease aircraft to enter their orders during the down cycle is to get a better position in the delivery queue. Dr. Benkard agreed that they would get both quicker delivery and a better price. Aircraft are priced through options contracts, with the price depending on the lead time of the order. An option exercised at 24 months has a better price than one exercised at 12 months.

Dr. Wessner noted the labor unions’ argument that Boeing weakens its efficiency when it reduces the labor force so quickly, and that it could better compete with Airbus by maintaining a more stable workforce. Dr. Benkard said he had

studied that question and had concluded that Boeing is actually following the most economical practice by ramping up during booms, producing as many planes as it can while orders are heavy, and laying off workers when the boom ends. This is because the inventory costs of large aircraft are so high that the company needs to move each unit out as quickly as it can, even if it means hiring too many workers and having to release some later.

Philip Auerswald asked about the primitives in the model from the standpoint of modeling the innovation process—especially the issue of sequential innovation leading to a technological crisis that needs to be resolved. The aircraft industry had a sequence of related products that were tracing out an industry-wide performance curve. Each of those products needed factory-level and product-level learning-by-doing, which would be a component of the price change. At the same time it had to deal with forgetting, and with the per-unit costs going up, as it went from one model to four models. He asked whether trying both to learn on existing products and to diversify over a range of products didn’t create a particularly severe learning and productivity challenge.

Dr. Benkard agreed. He added that he thinks that the story about “productivity waves,” or moving along a product frontier, also holds for the aircraft industry, but over a longer term. In 1956 the first aircraft with jet engines came out. In 1969 high-bypass jet engines were introduced and basically changed the industry. Today the technology is almost the same as it was in 1969, improved by smaller productivity increases. But he said he saw suggestions that additional, significant change is imminent. New technologies are usually tested out in the defense industry, where the technology is much higher, and some of these—including composite materials, fly-by-wire controls, and new wing designs—may trigger basic innovations in commercial aircraft in the next 10 years.

Dr. Mowery said that the negative effects of forgetting, or learning deficits, are important in semiconductors as well. Managers who are introducing a new process in a fab often find that much of the learning in yield improvement depends on engineering and managerial talent that is scarce and is spread thinly across multiple processes that all need to be debugged at the same time.

Dr. Auerswald suggested that the structural models discussed by Dr. Pakes form a framework for thinking about the evolution of an industry that has different firms, supply-and-demand questions, and people who are trying to predict the outcome of different investments. The modeler needs to get information from industry about what type of structure can be placed on that innovation process that makes sense. A limitless number of structures could be examined mathemati-

cally, but only a few can make sense for a staging of projects that go from one step to the next. In order to put the story together, he suggested, the model had best include the next generation of a discontinuous chain, incremental changes on a range of products, and new learning on an existing fixed product.

Mr. Morgenthaler noted his own involvement in the development and production of high-value components for jet engines from 1950 to 1957—the period of rapid buildup discussed by Dr. Benkard. He said that after it went to huge production, it ultimately made the transition to a huge spare-parts business. He noted that NASA is constantly developing new technology for spacecraft but lamented that the space agency is unproductive in promoting innovation for other industrial areas.

Dr. Benkard continued the discussion of spinoffs by saying that the airframe industry had a “parallel” industry, which was the defense industry, where much innovation is done. Fly-by-wire technology has already been spun off to commercial aircraft, which allows the use of airframes with advanced wing designs that would not otherwise fly commercially.

He also responded to Dr. Auerswald’s question about the learning curve. The industry starts out in a short-term mode to produce a certain amount of product and advance the learning curve. This is followed by a medium-term mode, concerned with product, how it is treated marginally, and perhaps the ability to bring in new products. He said that Dr. Auerswald was talking about either a longer-term investment process or an extreme outcome of the regular investment process. He said that that level had not yet been built into the models. So far, the models make use of the kinds of features with sufficient data, which are the types of investment processes that go on every year.

Ariel Pakes

Harvard University

Dr. Pakes said that in his summary he would offer more questions than answers.

He began with a question about the concept of coordinating investments by different firms. He noted that federal regulations are not clear about this subject. For example, the merger guidelines, which are designed to offer a legal framework on such issues, are essentially vacant on forms of coordination of anything

except pricing. One result from his experience with merger models is that the ability to coordinate the investment of both merging firms often generates an improvement for society. That is, if one asks the question of whether the new, multiproduct firm merged with full coordination is better for society, the answer is often “yes,” for a simple reason: A firm will invest to the point that the benefit of its investment equals the marginal cost of the investment. Part of that benefit is the gain of share from other firms.

A social planner does not count that share as a net increase. He concluded that some kinds of coordination, such as coordination of investment, might be in society’s interest, and that in this case the Department of Justice might not object to the practice. He said that although this area is not at the center of his expertise, he thought that the semiconductor industry was one of the few that had been successful in coordinating at least one kind of investment, namely, R&D investment.

Dr. Isaac agreed on the importance of this point and offered an example. From 1990 to 1995 he served as project manager for IBM’s 64-megabit DRAM development program, which was conducted jointly with Infineon and Toshiba. Of the two kinds of DRAM, the stack-capacitor type and the trench type, this collaboration was the only group working on the deep-trench type. Because the companies had considerable flexibility through subsidiaries and different formats, they were able to design joint research investments, joint manufacturing investments, and spinoffs. Today the deep-trench DRAM holds 40 percent of the world DRAM market. He said that neither IBM nor Infineon would have continued in the DRAM business past 1993 if they had not formed that alliance. “It enabled us to be productive in a manner that none of us could afford by ourselves,” said Dr. Isaac. “That kind of coordinated investment has been going on over this past decade to a very high degree and has indeed been central to the progress of the whole industry.”

Dr. Pakes raised a second set of questions regarding cyclicality. Most industries that sell durable goods, such as autos and airplanes, are cyclical. Some semiconductors go into durable goods, but he had not seen a demand analysis that takes this into account. One reason that a firm’s demand goes down tomorrow is that sales are heavy today. At a large auto manufacturer, the firm is aware that demand next year is linked to sales this year. He said he was surprised that the semiconductor industry did not anticipate a sales cycle in the same way.

Dr. Wessner cited recent signs that suggest global shifts in the market for semiconductors, with concentrated efforts by the Japanese to invest in wireless technologies and by the Europeans to focus on embedded appliances as well as wireless, based on standards that are unique to Europe. These trends, he said, may suggest that the market for semiconductors may fragment over the next few years.

Dr. Mowery pointed out that industries outside the durable-goods category also have cycles. He cited the paper industry and its “enormous, investment-driven cycles.” People invest in capacity in the good times, which are followed by down times. Over the past 15 to 20 years, he said, paper companies have consolidated, primarily through merger and acquisition.

He also mentioned the aluminum industry, which is characterized by “terrible” capacity and demand cycles. It also receives state support in the form of investment underwriting, which might be relevant to the semiconductor experience with the first years of SEMATECH, which was partly funded by government.

Finally, he said that the independent firms of the aircraft industry had responded to their inability to manage cycles by choosing strategies of merger, acquisition, or exit.

Dr. Pakes noted that mergers are the one practice virtually sure to draw the attention of the Department of Justice and Federal Trade Commission, largely because mergers have both price and demand implications for current products. He said that a merger is an “almost perfect coordination mechanism.” He restated his interest in better understanding the practice of the semiconductor industry in coordinating parts of its R&D work without merging. He wondered whether it might not be a model that could be applied in other industries, and whether it could be modified to apply to other forms of investment.

Dr. Wessner asked whether it was misleading to draw analogies between the semiconductor and airline industries, given their many differences. Dr. Mowery replied that there are both similarities and differences, with the differences most prominent at the technology level. He did see analogies in dealing with the problems of managing capacity and production in an industry facing wide swings in demand.

Dr. Pakes said that the reason for discussing aircraft was not so much to make a case for the similarities of the two industries. It was to explore the useful-

ness of modeling whole industries, and, given some years of experience in modeling the aircraft industry, to demonstrate the level of detail needed for productive modeling of the semiconductor industry.

Dr. Benkard closed this discussion by agreeing that the aspects that are central to the aircraft industry differ from those that are central to the semiconductor industry. He also refuted the point that semiconductor products reach the market faster than aircraft. He noted that experimental work with semiconductor products, such as chips with 8-nm separation, begins many years before a product reaches the market. Conversely, the first high-bypass jet engine was developed concurrently with the airframe it powered.

Dr. Pinto made the point that, in the semiconductor business, “transistors aren’t the product. The product is the chip that has 50 million transistors and the integrated circuit that goes with it.” The market lifetime of these products varies widely, from 9 months or less for the disk drive business to eight or ten years for an infrastructure chip.

Dr. Isaac seconded that point, and he reminded the group not to think of the semiconductor industry in monolithic terms, but to separate process technology, as seen in the foundry, from more integrated activities that bring products to market. He said that in the realm of process technology it may take 15 or more years to develop a useable new technology, such as copper, SOI (silicon on insulator), or silicon-germanium. Complex individual products may be somewhat quicker, such as IBM’s new Unix chip, which had been defined four years earlier. For discrete, specialized products, the tempo may be much more rapid; a chip for a storage drive may have to be designed, manufactured, and ready in nine months. In short, the industry is more diverse in terms of product lifetimes than many people realize.