Access Official Technical PDF

Share Summary:LinkedInEmail

Learning-by-doing in semiconductors creates a strategic tension: producing more today can lower a manufacturer’s future costs, but coordinated output restraint may limit those gains. An analysis of the DRAM cartel examines how this tension played out across overlapping product generations. Its findings suggest that collusion was more effective on an older generation than on newer ones, where faster learning made deviation more attractive.

Why learning curves matter to semiconductor competition

In DRAM manufacturing, experience gained through production can reduce the share of defective chips and lower marginal cost. The effect is tied to cumulative output: as a firm makes more chips, it learns how to manufacture them more efficiently. This is not a uniform advantage across a product’s life cycle. Engineers and production teams have generally had more time to solve process problems on established generations, so the scope for additional learning is lower than it is for a newly introduced generation.

That difference can affect firms’ incentives during a cartel. A company that increases production on a newer product may gain market share while also building experience that reduces its future costs. The greater the prospective cost reduction, the more appealing it can be to break an output agreement. In economic terms, the agreement must satisfy an incentive-compatibility constraint: the expected gains from continuing to cooperate need to outweigh the short- and long-term rewards of deviating.

Overlapping generations complicate output decisions

Semiconductor firms do not compete on one product alone. New DRAM generations arrive while older ones remain in production, and products differ in quality and cost. A firm’s output decision for one generation may therefore influence its position across the product portfolio. For example, production on a newer generation can affect the amount a firm chooses to make of an older one, while demand shifts across generations as customers adopt newer technology.

This multiproduct setting makes cartel behavior harder to interpret than a simple agreement to raise prices. Investigators need to consider learning, product life cycles, demand, and the relationship between generations—not just prices or aggregate production. The DRAM cartel analysis treated these dynamics as central to understanding differences in firms’ output choices.

What the DRAM cartel evidence indicates

The analysis compared cartel members with firms outside the agreement, using firm-level shipments across generations and time. It used a difference-in-differences approach: changes in member output during collusive periods were compared with changes among non-members, with estimates made separately for each generation. The model also accounted for cumulative output, demand associated with worldwide PC shipments, and cannibalization between generations.

The results were consistent with stronger output restraint on the older 16Mb generation than on the newer 64Mb and 128Mb generations. The study estimated a statistically significant relative reduction in cartel-member output on 16Mb, while the corresponding estimates for 64Mb and 128Mb were not distinguishable from those of non-members. Under the paper’s assumptions for translating member-level changes into total market effects, the estimated reduction in total 16Mb output was between 3.1% and 24.6%. That range reflects estimation uncertainty and the difficulty of separating member reductions from the response of non-cartel firms.

Price patterns provide additional context, though the analysis cautions that generation-level price data offered limited observations. During the cartel’s first phase, the 16Mb price rose as much as 25% above its pre-cartel starting value, while prices for newer generations fell. Price movement alone cannot establish how much output changed or why; the firm-level shipment analysis provides a more direct way to examine those decisions.

Testing for possible defection

One alternative explanation for weaker coordination on a new generation is that cartel members were deliberately building market share or cutting costs to compete with fringe firms. To assess whether firms may instead have deviated from a collusive arrangement, the paper examined the relationship between a firm’s order of entry into a generation and its subsequent output ranking.

In the DRAM data, earlier entry was generally associated with a higher later output ranking. The relationship was notably weaker for the 128Mb generation: its reported correlation was 0.43, below the correlations observed for the other generations in the sample. A weakened relationship may be consistent with later entrants overtaking firms that entered earlier, as could happen if a firm expanded output to capture learning benefits. This is evidence consistent with defection, not proof that a particular firm cheated or that every alternative explanation is ruled out.

Implications for antitrust analysis

The findings highlight why competition assessments in high-technology markets should account for dynamic production incentives. A useful investigation may need to examine:

  • Output by product generation: Aggregate shipments can conceal different patterns on older and newer products.
  • Cumulative production: Learning-by-doing can change the future cost of expanding output.
  • Entry timing and market-share rankings: Changes in the usual relationship between early entry and later position may help identify possible deviations.
  • Interactions across products: Cannibalization and demand shifts can shape output decisions across overlapping generations.
  • Cartel and fringe behavior: Non-participating firms provide a comparison for evaluating changes in member output.

These indicators should be interpreted together and in light of the limits of available evidence. The study’s framework is intended to help organize evidence for antitrust analysis, not to turn any single pattern into a definitive test for collusion.

Conclusion

Learning-by-doing can make coordination less stable when firms have more to gain from producing a newer generation and lowering their future costs. The DRAM evidence is consistent with stronger restraint on an older product and greater difficulty sustaining collusion on the newest generation examined. For automation and semiconductor decision-makers, the broader lesson is that production history and product life-cycle dynamics can influence competitive conduct. For competition authorities, analyzing generation-specific output, learning incentives, and the behavior of non-cartel firms can provide a more complete view of high-technology markets.

Access Official Technical PDF

About The Author