The Old Economics of the Storage Industry

Fri, 14 Aug 2026 21:05:00 +0800

When memory and SSD prices rise, the most convenient narrative is: supply has tightened, quotes have gone up, and storage manufacturers have finally become a good business. This picture is not untrue; it just captures the most exciting frame of the cycle.

The truly noteworthy part of the storage industry is the opposite: it has long involved high technical difficulty, massive manufacturing investment, and dense patents, yet it has repeatedly pushed companies into mergers, ownership changes, or exits. The question is not whether a price hike can happen, but whether a single price hike can change the economics across cycles.

First, separate the two. DRAM in a computer is runtime memory; the core storage medium of an SSD is NAND flash, along with a controller and firmware. The supply-demand dynamics and product cadence of the two are not entirely the same, but they share an uncomfortable underlying trait: their production lines and R&D are extremely capital-intensive, while unit costs can shift rapidly with process nodes, yield rates, and capacity utilization. When the market is strong, this machine amplifies profits; when supply loosens, it amplifies the pressure.

High Barriers to Entry: Why It Hasn’t Automatically Become a Good Industry

DRAM and NAND are certainly not low-tech products. Process technology, yield rates, equipment, intellectual property, and customer certification all eliminate participants who lack the capability for mass production. However, the threshold first determines who can take a seat at the table—it does not determine how the winnings are divided in each round.

When specifications become sufficiently standardized, buyers will still compare suppliers based on capacity, performance, delivery, and pricing. Established production lines bring heavy fixed costs: equipment doesn’t simply disappear when demand weakens, and R&D and manufacturing organizations can’t be scaled down at will on a quarterly basis. When supply is tight, scarcity drives prices up; when supply is loose, in order to ship products, amortize fixed costs, and preserve cash flow, price cuts become a choice that every company can understand.

This does not mean that every downturn is caused by the same decision, nor can “price war” be used as a single explanation for any one company’s collapse. What it describes is a structural tension: companies must invest in the next generation of technology and the next round of production capacity, and after many companies complete similar investments, the new supply will weaken current pricing power. High barriers have not eliminated competition; they have merely relocated it to a more expensive arena.

Patents are also often misunderstood as a wall that can permanently block competitors. They are closer to a ticket of entry, a bargaining chip, and a defensive tool. Without them, a company may struggle to manufacture or sell; with them, it still has to face the test of mass-production capability, scale, customer relationships, capital expenditure, and the balance sheet. Patents can raise the cost of entry and exit, but they cannot carry a company through a downturn.

Two Exiters, and One Company That Stayed Through Acquisition

Corporate history is more persuasive than abstract definitions. Elpida was a product of Japan’s DRAM industry consolidation: Hitachi’s and NEC’s DRAM operations merged in 1999 to later form Elpida; it went bankrupt and delisted in 2012, and was acquired by Micron in 2013. The merger was originally underpinned by expectations of scale and resource integration, but that path did not allow it to remain as an independent company in the end. For Micron, this was yet another acquisition to expand its manufacturing and product capabilities—not proof that it had never been through a down cycle.

Qimonda’s trajectory was even more acute. It was spun off from Infineon in 2006, and at the time possessed DRAM manufacturing capabilities, R&D personnel, and patent assets. Public records document that the industry supply glut and the decline in DRAM prices coincided with its major losses in 2008; in January 2009, Qimonda AG and its Dresden entity filed for bankruptcy protection. The most important point here is not to compress this history into “prices were low, so it went bankrupt,” but rather to see how the technology roadmap, fab investment, financing arrangements, and a market downturn all bore down on an independent manufacturer at the same time.

Path	Key Milestones	What It Illustrates
Elpida	Formed in 1999 through the merger of Hitachi’s and NEC’s DRAM businesses; went bankrupt and delisted in 2012; acquired by Micron in 2013	Consolidation can improve competitive standing but does not guarantee independent survival; downturns also reshape asset ownership
Qimonda	Spun off from Infineon in 2006; filed for bankruptcy protection in 2009	Technology, patents, and existing industry relationships cannot replace the cash, scale, and financing capabilities needed to endure cycles
Micron	Became one of the major memory manufacturers after acquiring Elpida	The path of survivors also includes absorbing assets during industry shakeouts, not just relying on a single technological advantage
Samsung, SK hynix, and Micron	Commonly cited today as the leading memory manufacturers	What the industry ultimately leaves behind is a small number of companies that can continuously invest, serve customers, and withstand cycles—not a set of names that naturally enjoy high profits
YMTC	Founded in 2016, focused on 3D NAND; began mass production of its first-generation product in 2018 and advanced subsequent iterations with the Xtacking architecture	New entrants must integrate design, mass production, and end products together to qualify for NAND cost and supply competition
CXMT	Founded in 2016, focused on DRAM; announced mass production and sales of 8Gb DDR4 in 2019	New entrants in DRAM face not a single product launch but a continuous test of yield, customer validation, capacity ramp-up, and cycles

This table is not meant to rank companies. It simply lays out something that tends to be obscured by pricing news: storage companies must simultaneously address four challenges—technology, scale, customers, and capital. Losing support on any one of these during a downturn can turn “having the technology” into an insufficient condition.

China Storage Is Not One Company, But Two Different Uphill Roads

If the preceding text only lists Samsung, SK hynix, and Micron, it portrays today’s industry landscape as too static. Both China’s Yangtze Memory Technologies (YMTC) and ChangXin Memory Technologies (CXMT) should be included in this picture, but they should not be combined into one generalized “domestic storage” story.

YMTC follows the NAND route. According to company information, it was established in Wuhan in 2016, focusing on 3D NAND; its first-generation products began mass production in 2018, followed by product iterations centered on the Xtacking architecture. It has product arrangements ranging from wafers to embedded storage, consumer-grade and enterprise-grade SSDs, which precisely illustrates that NAND competition is not limited to a single flash memory wafer: it also requires integrating controller, packaging, product form factors with customer use cases.

CXMT follows the DRAM route. According to its company timeline, CXMT was founded in 2016, announced mass production of 8Gb DDR4 in 2019, and initiated product sales in the same year. This milestone is not proof that it has already “resolved the DRAM cycle”; rather, it marks the starting point at which a new manufacturer truly enters the market. What follows still requires repeated validation in terms of yield rates, costs, product generations, customer qualification, and the pace of capacity expansion.

Therefore, the significance of these two companies for the industry’s history goes beyond simply adding two more suppliers. They place old questions in front of new players: whether high investment can be transformed into stable, marketable output, whether the technological roadmap can sustain continuous iteration, whether customers are willing to complete long-term validation, and whether capacity expansion during boom periods will not turn around and erode returns during downturns. The progress of Yangtze Memory Technologies and ChangXin Memory will reshape the competitive landscape, but it will not free memory manufacturing from these constraints.

How Technical Roadmaps Push Companies Toward Different Fates

Memory is not a single technology. DRAM and NAND differ in physical structure, product generations, and patent portfolios, so a company’s history cannot be read simply by “who is bigger.”

DRAM: The Survivor Isn’t Always Leading, But Doesn’t Break Through Generations

Samsung’s roadmap has been one of introduction, catching up, and then turning generational R&D into manufacturing scale. In 1974, the Samsung Group entered the semiconductor business by acquiring Korea Semiconductor; in 1983 it announced its entry into DRAM, and subsequently produced 64Kb DRAM. The capability was not built out of thin air in its early days—public historical records note that its DRAM development relied on Micron technology, and its SRAM and ROM on Sharp technology. This origin clearly illustrates what the “patent wall” really means: a laggard can enter the race through licensing, introduction, and engineering digestion, but after entering, it must still continuously deliver next-generation devices, processes, and mass-production capabilities. Samsung subsequently pushed forward 64Mb, 256Mb, and 1Gb DRAM in succession during the 1990s, and brought V-NAND into commercial products in 2013. Its strength is not that any single patent remains perpetually exclusive, but rather that DRAM, NAND, packaging, end products, and capital expenditure can relay from one another across many product generations.

The history of SK hynix is a story of restructuring aimed at “surviving first.” Its predecessor, Hyundai Electronics, was founded in 1983. It initially entered the market with SRAM but shifted to licensing DRAM designs due to poor yields, then accumulated manufacturing expertise through OEM production; in 1998, it absorbed LG Semicon. During the industry downturn in 2001, Hynix suffered severe losses, subsequently divesting non-core businesses and separating from the Hyundai Group, before being acquired by the SK Group in 2012. Today, when people view the company through the lens of HBM, it is easy to overlook this long journey: a technology roadmap is not merely a cell structure diagram, but an organizational arrangement that keeps R&D, equipment, talent, and customer supply running even during industry downturns.

Micron is another example from the U.S. route. It started out as a design consulting company in 1978 and completed its first wafer fab in 1981 to produce 64K DRAM. Afterwards, it engaged in both device and process R&D while expanding its boundaries through partnerships and acquisitions—for example, establishing IM Flash Technologies with Intel for NAND, and acquiring Elpida in 2013. Therefore, Micron’s “patent route” should not be understood solely in terms of litigation or licensing: process capabilities such as high-k dielectric thin films and double patterning accumulate into patents and manufacturing expertise, while the acquisition of Elpida incorporated Japanese manufacturing, customer, and product assets into the new corporate entity. The patent portfolio helps the company negotiate and defend itself, but whether an acquisition succeeds still depends on whether the factories, personnel, and customer relationships can be integrated.

Elpida and Qimonda both illustrate that knowing how to make DRAM does not mean a company will survive to the next generation. Elpida was formed in 1999 through the merger of Hitachi’s and NEC’s DRAM businesses, and later absorbed Mitsubishi’s DRAM operations as well; it went bankrupt in 2012 and was acquired by Micron in 2013. Qimonda was spun off from Infineon in 2006, bringing along its manufacturing, R&D, and patent assets, but the oversupply and price collapse of 2008, combined with major losses, drove it to file for insolvency protection in 2009. Their outcomes can be read as the same underlying constraint: patents and existing fab lines only earn a company a ticket to the next round—they do not pay the bills for next-generation process nodes, yield ramp-up, or financing through a downcycle.

NAND: From “Scaling the Plane” to “Splitting the Process and Stacking It Up”

The NAND competition more clearly reveals how technology roadmaps shift the focus of patents. In the early stages, the emphasis was on fitting more cells into a smaller area. As planar scaling advanced, cell-to-cell interference, reliability, and cost-per-bit became increasingly difficult to control simultaneously. The industry then turned to 3D NAND: cells are stacked vertically, and the competition shifts from planar dimensions to etch depth, stacking architecture, peripheral circuits, bonding, yield, and controller co-design.

Toshiba-related assets are one of the early sources along this route. Toshiba engineer Fujio Masuoka invented flash memory in 1980; Toshiba’s memory business was spun off as Toshiba Memory in 2018 and renamed Kioxia in 2019. The reason for the company name change lies not in the technology itself, but in the parent company’s capital and operational pressures; this serves as a reminder that having a long-standing NAND technology heritage does not automatically shield memory assets from corporate-level financing and governance issues. The joint manufacturing relationship between Kioxia and Western Digital demonstrates that NAND patents, fabs, and customers are not always housed within the same company.

Samsung’s NAND path is to bring its continued capability in memory manufacturing into 3D stacking. Micron, beyond its NAND partnership with Intel, continues to develop its own 3D NAND products. Both are answering the same question: not whether a multi-layer structure can be produced in the lab, but whether such a structure can be stably etched, filled, tested, and—after passing through SSD controllers, firmware, and enterprise-grade validation—turned into a sustainably sellable product.

YMTC has placed this NAND roadmap in a Chinese context. The company was established in 2016, and its first-generation 3D NAND entered mass production in 2018. The company continues to cover wafers, embedded storage, and consumer- and enterprise-grade SSDs, completing the chain from bare die to deliverable products.

Two Different Entry Paths

YMTC’s origins cannot be understood from a single architecture alone. When YMTC was established, it was built on the foundation of Wuhan Xinxin Semiconductor Manufacturing (Wuhan Xinxin), which became its wholly-owned subsidiary; the company explicitly used Wuhan Xinxin’s existing 12-inch R&D and manufacturing capabilities as the basis for developing large-scale memory. Earlier, Wuhan Xinxin had already formed a joint R&D team with Cypress (formerly Spansion) to develop 3D NAND, and the two parties announced the signing of a development and cross-licensing agreement. Its initial accumulation came from existing manufacturing and R&D platforms, cooperative development, and licensing, rather than from acquiring complete technology systems through the purchase of overseas NAND manufacturers. Xtacking is an architectural innovation that can be specifically pointed out atop this accumulation: it did not eliminate existing rights and engineering constraints, but reframed the competition around how the array, peripheral circuits, and bonding are collaboratively optimized.

ChangXin Memory faces a different kind of test. It was established in 2016 as a DRAM company, broke ground on the Hefei Fab A in 2017, completed the 8Gb DDR4 tape-out in 2018, and in 2019 finished testing, announced mass production, and secured its first orders. This timeline breaks “having a DRAM solution” into more realistic steps: build the fab first, then tape out, then test, mass-produce, and deliver. The subsequent product generations, yield rates, and module and end-customer validation will still determine whether it can turn manufacturing capability into sustained commercial capability.

ChangXin Memory cannot be described as having “acquired” Qimonda either. A joint announcement from 2019 shows that WiLAN’s subsidiary Polaris signed a patent licensing agreement with ChangXin Memory, along with a separate patent acquisition agreement: Polaris licensed ChangXin a number of patents related to DRAM technology developed by Qimonda, while ChangXin agreed to acquire a portion of Polaris’s DRAM patents. The announcement uses the phrase “agreed to acquire,” so this only indicates that the two parties signed a licensing agreement and agreed to an acquisition—it cannot be written as ChangXin having already completed the patent assignment; nor can it be elevated into an acquisition of Qimonda, or an assumption of its fabs, personnel, customers, and manufacturing know-how. ChangXin subsequently also signed a broad DRAM patent licensing agreement with Rambus. These arrangements lowered certain intellectual property barriers to entry, but they do not replace the accumulation of DRAM process technology, yield, and customer qualifications. The currently available public materials are not sufficient to point to a specific, independently verified DRAM device or process invention by ChangXin in the way that Xtacking can be explained; the more prudent conclusion is that its visibly public starting path simultaneously includes patent licensing, an agreed acquisition of some patents, as well as the independently completed manufacturing ramp of building its fab, tape-out, testing, and mass production.

Grouping YMTC (Yangtze Memory Technologies Corp) and CXMT (ChangXin Memory Technologies) together as “domestic memory” obscures a critical distinction: the former primarily focuses on the stacking and integration of 3D NAND and system products, and can point to an architectural path via Xtacking; the latter primarily focuses on the continuous scaling of DRAM cells, processes, and memory products, and its publicly verifiable intellectual property path is more directly expressed through licensing and agreed-upon acquisition of certain patents. Neither path represents a one-time breach of the patent wall; rather, each secures certain entry conditions and continues to face manufacturing and market scrutiny.

What Exactly Does the Patent Wall Block

A patent wall is not a wall with only “yes” or “no”, but has at least three layers.

The first layer is interface and standard-related rights. Interfaces such as DDR, LPDDR, PCIe, and NVMe enable systems to interoperate and allow customers to compare substitutions among qualified suppliers. Standards reduce the friction of system integration, yet they do not give every chip vendor the same manufacturing capabilities; patent licensing, standards organizations, and potential disputes therefore persist over the long term.

The second layer is device and process patents. DRAM capacitors, transistors, word lines, and patterning schemes; NAND charge-storage structures, inter-layer etching, stacking, bonding, peripheral circuits, and reliability schemes may all be covered by long-standing layouts from different companies. This layer of rights explains why the transition from early technology introduction to V-NAND at Samsung, from process R&D to NAND partnerships at Micron, and from array/peripheral-circuit separation to bonding at YMTC, is not simply a matter of switching to a new product roadmap. A new roadmap must both avoid stepping into existing rights and accumulate its own protectable inventions.

The third layer often does not exist in patent texts, but rather in manufacturing know-how and customer qualification. Which parameters need to be tuned during yield ramp-up, which process step defects originate from, how the controller handles NAND lifespan and bad blocks, and how enterprise SSDs pass long-term consistency tests — some of this knowledge is covered by patents, while some exists as experience that is difficult to transfer. CXMT’s timeline from tape-out to orders demonstrates the existence of customer qualification: producing the chip does not mean the product has already been accepted by the procurement system.

This is exactly why patents are both powerful and not powerful enough. When Qimonda exited, its intellectual property and technological assets still existed, yet the company itself did not continue to exist as a large-scale DRAM manufacturer as a result. Conversely, new entrants hold patent applications and architectural innovations, but still need to turn them into repeatable manufacturing and customer trust. Cross-licensing among established manufacturers does not mean competition disappears; rather, it acknowledges that no single company can easily circumvent the entire accumulated history. Real price competition will continue to occur among those who have already secured their entry ticket.

HBM pushes the problem into a more complex position. It is not a simple price increase on ordinary DRAM chips; instead, it ties memory dies, stacking, packaging, testing, and customer-side validation with computing clients more tightly together. The patents, processes, and supply chain coordination here may raise switching costs for a period of time, but it still has to pass the same historical test: will other manufacturers catch up, will the new capital lead to oversupply, and will customers accept new alternatives. What AI changes is the location where competition happens; it does not automatically eliminate competition.

Concentration is Not a Guarantee of Profits

When an industry consolidates down to a few players, it’s easy to equate concentration with stable profits. History offers an uncomfortable footnote: the European Commission in 2010 fined DRAM price cartel participants a combined €331 million.

This is not to say that a small number of vendors are necessarily breaking the law, nor does it use a single regulatory case to predict future pricing. It only illustrates that a small number of vendors, in itself, cannot replace real supply and demand, nor can it be taken as evidence that profits can be naturally coordinated. The returns of the storage industry also depend on how capacity expands, whether customers can switch suppliers, whether products are substitutable, and whether newly added capital can leave sufficient cash flow before the next downturn arrives.

AI Has Changed the Problem, Yet Hasn’t Logged Out of History

The HBM demand brought about by AI is indeed a new variable. It reminds us, at minimum, that we cannot treat all memory products as the same commodity, nor can we mechanically apply every supply-demand shift in traditional DRAM to all product segments.

However, this article does not use current-period HBM supply, pricing, capacity, or profit data, so it cannot be used to judge how long this period of high profitability will last. A more prudent statement would be: if a certain type of product experiences stronger constraints in customer qualification, supply capability, or portfolio mix, it may, for a period of time, change the degree of shortage and the distribution of profitability; to prove that it has altered the returns across the cycle, it still needs to go through more rounds of capital expenditure, competitive catch-up, and demand fluctuations.

So, in the face of rising memory and SSD prices, rather than asking “how high can prices go this time,” it is better to ask a few slower questions: whether capital expenditure during the boom will accelerate again; how easily the product can be substituted on the customer side; whether the new profits can be retained as cash rather than turning into the next round of excess capacity; and whether the company has a strong enough balance sheet to make it through to after the next downturn cycle.

The storage industry has never been one that cannot make money. It tends to make plenty of money when supply is tight. History keeps reminding us of something else though: the most dazzling quarter is usually the worst sample to use in place of the industry’s entire history.

This article does not constitute investment advice.

References

Micron Memory Japan (formerly Elpida) Corporate History Summary, verified on 2026-08-14.
Qimonda Entry and Company Restructuring, Bankruptcy Protection Records, verified on 2026-08-14.
SK hynix Company Summary, verified on 2026-08-14.
Samsung Electronics: Semiconductor and Memory Technology Development Records, verified on 2026-08-14.
SK hynix: Hyundai Electronics, Hynix, and SK Restructuring History, verified on 2026-08-14.
Micron: Founding, NAND Partnerships, and Elpida Acquisition Records, verified on 2026-08-14.
Kioxia: Spinoff and Renaming of Toshiba’s Memory Business, verified on 2026-08-14.
YMTC: Company Introduction and Development History, verified on 2026-08-14.
YMTC: Xtacking Architecture Explanation, verified on 2026-08-14.
CXMT: Company Introduction and Development History, verified on 2026-08-14.
YMTC Establishment Announcement: Built on the Foundation of Wuhan XMC, verified on 2026-08-31.
Wuhan XMC: Joint 3D NAND Development with Cypress (formerly Spansion), verified on 2026-08-31.
Spansion and Wuhan XMC: Joint Development and Cross-Licensing Agreement Announcement, verified on 2026-08-31.
WiLAN/CXMT: Joint Announcement on DRAM Patent Licensing and Acquisition Agreement, verified on 2026-08-31.
CXMT and Rambus: DRAM Patent Licensing Announcement, verified on 2026-08-31.
European Commission: Penalty Announcement Against DRAM Producers for Price Cartel, 2010-05-19.

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.

CXMT on Uncle Xiang's Notebook