Uncle Xiang's Notebook Collecting Moments, Marking Time
← Article list

The Old Economics of the Memory Industry

From Elpida and Qimonda to the three major memory makers: why a favorable pricing window rarely rewrites the cycle

Codex (GPT-5) v8.11.1

Rising memory and SSD prices make the profitable side of the memory business easy to see. These products are almost unavoidable in modern computing, the technical barriers are high, and new factories are expensive. Manufacturers might therefore seem entitled to a good business for years to come.

Yet bankruptcies, sales and restructurings are just as prominent in its history. Elpida and Qimonda both had technology, factories and patents; neither survived as the same independent competitor. The question is not whether memory can make money. It is whether profits in a good market can pay for the next technology generation, capacity expansion and the next downturn.

On this page

High barriers decide who can enter, not how much entrants earn

First, separate two products. DRAM is working memory; NAND is a major storage medium in products such as SSDs, which also contain controllers, firmware and other components. Their technologies, supply and demand, and product cycles differ. Calling both “memory” does not justify explaining them with one price curve.

They share heavy research and manufacturing commitments. New processes require continuing investment, installed production lines create fixed costs, and yields and utilization affect unit costs. Where specifications and qualification allow substitution, buyers still compare prices, performance and delivery.

When prices fall, each company wants to ship more to spread its existing costs. Falling behind in the next technology generation can also leave it with higher costs later. Individually rational survival decisions do not necessarily add up to an industry with high returns. Technical barriers restrict competition to a few capable companies while making every round more expensive.

This is not a universal explanation for every failure. Debt, exchange rates, customer mix, parent-company conditions and technology choices play different roles. History helps show how these pressures meet; “price wars” cannot substitute for an explanation of each company’s fate.

Companies exit; technology and capacity may remain

Elpida originated in the consolidation of Hitachi’s and NEC’s DRAM businesses in 1999 and later incorporated related Mitsubishi operations. It filed for bankruptcy in 2012, and Micron completed its acquisition in 2013. Qimonda was spun out of Infineon in 2006 and filed for bankruptcy in 2009.

Their histories remind us that scale, processes and intellectual property do not replace cash flow. In a downturn, a company must fund more than current production: financing, equipment and next-generation research still come due.

But a manufacturer’s collapse does not mean its supply capability disappears permanently. Factories, patents, customers and employees can transfer to another owner through acquisition or restructuring. The disappearance of an old name may reduce the number of competitors while giving survivors a more complete set of assets. Counting corporate names alone does not measure how much capacity has actually left the industry.

Company or assets Key events What they suggest about the industry’s economics
Elpida Consolidated Japanese DRAM operations; bankruptcy in 2012, acquisition by Micron in 2013 Consolidation does not eliminate cycles; acquired assets may continue producing
Qimonda Spun out in 2006; bankruptcy in 2009 Patents and factories do not guarantee enough funding for the next round
Micron Built manufacturing capabilities while extending its products and assets through partnerships and acquisitions Survival requires integration and financing as well as research

The three DRAM survivors relied on more than one technology

Samsung entered DRAM in 1983, repeatedly connecting successive generations of research to manufacturing at scale, while expanding into NAND and packaging. Its advantage is better understood as sustained execution across research, mass production and investment than as the possession of one decisive early patent.

SK hynix’s predecessor, Hyundai Electronics, was founded in 1983. It integrated LG Semicon in 1999, later underwent asset disposals and organizational restructuring in difficult years, and joined SK Group in 2012. Looking back from its position in HBM today, it is easy to overlook how much capability lies in simply keeping operations intact through a downturn. Research teams, factories and customer supply relationships cannot be rebuilt instantly when prices recover.

Micron was founded in 1978 and began production at its first wafer fab in 1981. Alongside its own device and manufacturing capabilities, its NAND partnership with Intel and acquisitions such as Elpida added scale, customers and products. An acquisition transfers assets, but successful integration still requires manufacturing, people and customer relationships to work together.

None of these histories proves that a company will lead forever. Together, they show that leadership in memory is not a permanent reward for one invention. It requires investment, delivery and customer acceptance across successive technology cycles.

NAND moved competition from a flat surface to the whole process

Early NAND mainly increased density by shrinking planar structures. Eventually, interference between cells, reliability and costs became harder to reconcile. With 3D NAND, cells are stacked vertically, shifting more of the competition toward deep-hole etching, structures between layers, peripheral circuits, bonding and yields.

A higher layer count is visible, but it is not synonymous with lower costs or reliable delivery. A structure must be manufactured consistently; the resulting dies must work with controllers, firmware and customers’ operating conditions. Enterprise SSDs in particular require lengthy validation of reliability and consistency.

Toshiba separated its memory business in April 2017, completed its sale to a Bain-led consortium in 2018, and renamed Toshiba Memory as Kioxia in 2019. Separation, sale and renaming were three distinct events. They did not mean that the technological tradition ended, but they show how a parent company’s finances and operations can alter the ownership of memory assets. After Sandisk’s separation from Western Digital in 2025, it was Sandisk that renewed the joint-manufacturing agreement with Kioxia in 2026. Joint manufacturing is another reminder that brands, patents, factories and customers need not all sit within one company.

Samsung, Micron and other NAND producers face the same engineering task: producing more bits consistently and selling them in suitable products. Layer counts announced at a technology launch cannot establish who will earn lasting profits.

YMTC and CXMT took different paths into the industry

YMTC focuses on NAND; CXMT on DRAM. Treating them as one story of Chinese memory breakthroughs obscures the engineering and intellectual-property problems each must solve.

YMTC was founded in 2016 and mass-produced its first-generation 3D NAND in 2018. Its beginnings were connected to XMC’s manufacturing and research base in Wuhan. XMC’s earlier joint development and licensing relationships with Cypress and the former Spansion were part of that accumulated foundation. YMTC did not acquire an overseas NAND company and inherit a complete, ready-made industrial capability.

Xtacking offers a concrete architectural direction: coordinating the separate manufacture and optimization of memory arrays and peripheral circuits with their bonding. Such a route can build technology and patents of its own, but it does not remove other process requirements, existing rights or customer qualification. The product chain from wafers to embedded storage and SSDs still has to be completed.

CXMT was also founded in 2016. Construction of Hefei Fab A began in 2017, an 8Gb DDR4 design was taped out in 2018, and testing, mass production and sales progressed in 2019. The sequence matters: a design, samples, volume production and customer orders are different stages. Reaching one does not mean every capability is already in place.

Its intellectual-property route also needs precise wording. The 2019 announcement by Polaris and CXMT covered licensing DRAM-related patents developed by Qimonda, alongside a separate agreement to acquire some patents. Obtaining a license, agreeing to buy patents and acquiring Qimonda itself are three different things. The phrase “agreed to acquire” does not establish that the transfer had closed, much less that factories, employees, customers and manufacturing know-how moved with it. Later licensing from Rambus likewise addressed certain rights; it did not replace CXMT’s own manufacturing ramp.

New entrants change supply and competition without escaping the industry’s constraints. After demonstrating that a product can be made, they still need to address costs, yields, product generations, customer qualification and funding through a downturn.

Patent barriers are powerful, but they are not the whole barrier

It helps to distinguish three layers.

Layer What it constrains What it cannot replace
Interfaces, standards and related rights How products interoperate with systems, and which rights must be addressed Manufacturing capability and volume-production costs
Device and process patents Cell structures, patterning, stacking, bonding and other technical solutions Stable yields and continuing engineering improvements
Manufacturing know-how and customer qualification Defect handling, reliability, firmware compatibility and procurement approval Continuing investment and market demand

The last layer is not fully written into patents. A license can settle whether a company is entitled to use a technology without settling whether it can manufacture it reliably. Owning a factory does not mean customers will put its products into their systems.

Patents are therefore both barriers and tools for negotiation and cooperation. Established manufacturers can cross-license and still compete. An exited company’s patents can retain value even when that company can no longer sustain manufacturing. Treating intellectual property as a guarantee of permanent profits misreads both survivors and failures.

Fewer suppliers and AI demand still need a full-cycle test

Greater concentration can change competitive conditions, but it does not establish that profits can automatically be coordinated. In 2010, the European Commission imposed approximately €331 million in fines over a DRAM price cartel. That historical case does not predict today’s prices, but it is reason enough not to treat a small supplier count as a promise of lasting profits.

HBM adds new constraints. It ties DRAM dies more closely to stacking, advanced packaging, testing and qualification by computing customers. It is not simply conventional memory. Some bottlenecks can slow supply responses and raise switching costs; they also attract capital seeking to catch up.

To establish that AI demand has changed the industry’s returns across cycles, we still need to see when competitors catch up, how much usable supply new investment creates, whether demand persists, and whether boom-time profits leave cash behind or capacity that must be depreciated in the next cycle. One highly profitable quarter cannot answer these questions.

Memory is not an industry that never makes money. Its problem is that the good times make it easy to forget the bill already taking shape for the next round of competition. A quarter may be enough to observe rising prices. To judge the quality of the business, peaks and troughs belong in the same accounts.

References

写作附记

Original Prompt

$blog-writer SSDs and memory have recently surged in price, but why do people say that the semiconductor storage industry is not a good one? Starting from price wars, patent walls, the development paths of the surviving companies, and the experiences of semiconductor manufacturers that have already gone bankrupt, let’s explain the past of this industry. It’s unclear whether the high profits brought by AI can be sustained; let’s first sort through what the history looks like. Rewrite the articles from these three months. Note: if an article was not written by AI, do not rewrite it.

Table of contents

Related content

01
  1. 01

    Capacity Expansion Is Underway. Why Won't Memory Get Cheaper Soon?

Comments

Comments will load when you scroll here.