A 60-Year History of Memory Semiconductors

01Why Do Only Memory Chip Prices Ride Such a Roller Coaster?

Semiconductors are broadly divided into two worlds: system semiconductors (logic chips) and memory semiconductors. System chips, such as CPUs and GPUs, compete on design. Circuit architectures differ from company to company, and complex software is built in layers on top of them. Once a chip establishes its place, customers cannot easily switch to another product. In that arena, making something uniquely different is your greatest weapon.

Memory is the exact opposite. DRAM follows strict, standardized specifications: chips from Company A and Company B plug into the very same slot and communicate using the exact same electrical signals. This standardization is essential so computer makers can swap parts freely. Yet while convenient for buyers, it is a brutal rule for manufacturers, leaving virtually no room to stand out through design alone. How do you win, then? The only battleground left is production cost.

That is why memory is an industry where economies of scale directly decide the winner. Whoever yields more intact, defect-free chips from a single round silicon wafer wins the game. If you pattern circuits a little more densely, you get more chips per wafer; if you trim defect rates a bit further, you have more chips you can actually sell. Accomplish both before your rivals, and even when selling at the exact same market price, you are the only one left with a profit.

Here lies the catch: to acquire that capability, you must build the manufacturing plant first. Building a single state-of-the-art fab is known to cost tens of trillions of won (billions of dollars), and that investment decision must be made long before product prices rise. A fab takes years to build, yet the decision to build is driven by the immediate prices right before your eyes.

This gives rise to the industry's signature rhythm. When prices are high, every manufacturer rushes to expand capacity at once. Around the time those new fabs begin running simultaneously, supply floods the market and prices crash. As prices collapse, everyone halts investment together. A few years later, a shortage returns, sending prices soaring again—triggering yet another wave of joint expansion. The industry calls this wild swing the silicon cycle. In essence, the business cycle is etched entirely into a single industry.

In this game, financial stamina to survive a downturn is just as vital as technological prowess. The company that holds out and keeps investing when prices scrape bottom becomes the ruler of the next boom. Conversely, stumble once, and you rarely get a chance to recover, simply because you will have no running fabs ready when the next boom arrives.

In the story ahead, the crown changes hands three times: from the United States to Japan, from Japan to South Korea, and now it is trembling once again. All three shifts unfolded strictly atop these very rules.

Worth remembering Only by grasping the rules of this game first can you see that the three changes in market leadership were never an accident.

02How to Remember 1 Bit with a Single Transistor

In 1966, Robert Dennard at IBM Watson Research Center was examining memory circuits. At the time, storing just 1 bit required six whole transistors. It worked by having six transistors support one another to hold a state. It was reliable and fast, but it took up far too much space. To expand memory capacity, the cost multiplied right along with it.

Dennard had a different idea: wouldn't one transistor and one capacitor be enough? The capacitor could act as a tiny bowl holding a small charge of electricity, while the transistor simply served as a gate opening and closing that bowl. If electricity was stored inside, it was a 1; if empty, a 0. To read it, you opened the gate to look inside; to write, you opened the gate to fill or empty it. In 1968, he was granted a U.S. patent for this structure.

The catch was that electricity gradually leaked out of the bowl. This meant you had to constantly read and recharge it before the contents blurred away. Because of this continuous recharging, or refreshing, the word "dynamic" was attached, giving birth to DRAM. It was a cumbersome mechanism, but it slashed the component count from six down to two. It meant that you could fit far more cells into the same area. In the world of memory, this was not a mere refinement—it was a breakthrough that dropped costs by orders of magnitude.

Yet here lies the story's first twist. In 1970, Intel launched the 1103, making it the very first commercially successful DRAM, but the 1103 did not actually use Dennard's 1-transistor cell. Instead, it used a 3-transistor design that was easier to build. It stored 1,024 bits, came in an 18-pin package, and cost about $60. Intel chose not the elegant design, but the one that delivered viable yields—the share of functioning chips that existing manufacturing processes could reliably produce.

Even so, the 1103 displaced magnetic-core memory, which had ruled computer memory until then. Magnetic cores were crafted by hand, threading fine wires through tiny magnetic rings smaller than grains of rice. Increasing capacity meant needing that many more human hands. A slice of silicon, on the other hand, could be mass-produced in multiples with the exact same effort once the process was nailed down. It was the moment memory woven by hand turned into a product stamped out in a factory.

In the semiconductor world, "technology that works" is entirely different from "technology you can manufacture." The signal that the battle would be decided by manufacturing rather than sheer invention was already plain to see in the very opening year of the industry.

Worth remembering The rule of this industry—that the inventor does not necessarily take the spoils—reveals itself here for the very first time.

03Why Did Intel Abandon the DRAM It Invented?

In the 1970s, Intel was a memory company. Today it is known as a CPU maker, but for a long time, DRAM was the product that kept the company alive. There was even a time when its DRAM market share exceeded 80%. But by 1984, it had dropped to 2~3%. It all happened in little more than a decade.

What happened? Japanese makers began churning out DRAM with far fewer defects in massive volumes. Customers building mainframe computers took one look at the failure rates and switched suppliers, leaving Intel outmatched in both volume and yield. True to the rule from the previous chapter, this game was decided by manufacturing prowess, not design. Intel was standing on a battlefield where the gift of drawing clever circuits was of no use.

What made things even more difficult was DRAM's status inside the company. The top research talent and the finest production lines were all tied to DRAM. The money was already coming from microprocessors, but the company's body was still built in the shape of a memory maker.

In 1985, Andy Grove asked Gordon Moore what a new CEO would do if the board kicked them out. The answer was obvious: he would get out of DRAM. Then Grove asked, why shouldn't we just do it ourselves? That day, the two men made the decision to exit the market their own company had created.

The withdrawal was completed in 1986, and that year Intel posted a loss of $173 million. Shutting down fabs and letting people go meant cutting away the company's very roots. Grove later called this decision the best strategic choice of his career.

Where Intel went next was microprocessors. This was not a product where someone else set the specifications, but one where Intel defined them itself. And as more software piled up on top of it, customers found it harder and harder to leave. Once software was written for Intel's architecture, switching to another company's chip rendered all those programs useless. Intel had not merely run away from a price war—it had switched to a game played by rules it could win.

This decision was remarkable because, across the rest of our story, almost no other company managed to pull off the same move. Most clung to what they were once good at, waiting for prices to rebound. And many of them never lived to see that recovery.

This was how most major American corporations walked away from DRAM. Only Micron, then just getting started in Idaho, remained behind, and leadership of the market passed to Japanese companies. For Intel, the choice was the right one. But as the price of that right decision, the entire memory industry crossed the Pacific.

Worth remembering A rare example of switching the board entirely when the winning formula stops working—a choice most companies that follow will fail to make.

04Japan Won on Quality, but What Tripped It Up?

In 1976, Japan's Ministry of International Trade and Industry (MITI) formed the VLSI Technology Research Association. Five companies—Fujitsu, Hitachi, Mitsubishi Electric, NEC, and Toshiba—set aside their competition for a moment and sent researchers to a single lab. The idea was to compete on end products, but jointly research the equipment and manufacturing processes used to make them. They pooled their efforts only on the parts that were too expensive for any single company to handle alone.

The association operated until 1980. From 1976 to 79, about $150 million in government subsidies went in, producing more than 1,000 patents. Results followed quickly. In 1977, NTT announced that it had developed the world's first 64K DRAM. In the early 1980s, Japan held over 70% of the global 64K DRAM market. In 1988, Japan accounted for about 51% of the global semiconductor market, while the United States held about 37%.

How Japan won wasn't about price; it was about quality. Back then, the big customers for DRAM were mainframes used by corporations and research labs. A single mainframe packed thousands of memory chips, and if just one went bad, the entire machine ground to a halt. Consequently, a tiny difference in defect rates decided purchasing choices. Japanese makers aligned their entire organizations around tightly controlled processes to eliminate defects, and with that approach, they overwhelmed American manufacturers.

DRAM wasn't all Japan created. In 1987, Fujio Masuoka's team at Toshiba presented NAND flash at an academic conference. It was memory that retained its data even when powered off—the technology that would later become the foundation for SSDs and smartphone storage. Yet in 2002, the company that took the top spot in that market by becoming the first in the world to mass-produce 1Gb NAND was Samsung Electronics. The rule that the inventor doesn't necessarily take the spoils played out once again.

Meanwhile, the demand for DRAM was also shifting—from mainframes to personal computers. A person buying a PC doesn't worry about ten years down the road. It's a device replaced after a few years, so being cheap mattered more than lasting a long time. Processes and organizations optimized for high-reliability, high-price products were too heavy in the face of products that were good enough and far cheaper.

Being "heavy" didn't mean they were lazy. In fact, it was quite the opposite. The more rigorous the inspection stages and process controls built to reduce defects, the more a decision to strip them away to cut costs felt like denying the company's pride. The very formula that brought you victory is always the hardest thing to let go.

Japan's market share peaked in 1988 and began draining away in earnest starting in 1992. People often credit this to trade pressure from the United States, but that pressure was merely the trigger. The real reason was that the nature of the products being sold had changed in the meantime, and Japan failed to transform itself to match that shift.

Worth remembering Here, it becomes clear for the first time: the very formula that brought victory becomes a burden in the next round.

05Why Did Regulations Meant to Stop Japan End Up Fueling Korea?

On September 2, 1986, the U.S.–Japan Semiconductor Agreement was signed. Japan promised to stop dumping and open its market. A secret side letter even included a target for foreign chips to take 20% of the Japanese market within 5 years. Rather than imposing tariffs, it directly meddled with prices and market shares—an intervention close to a non-tariff barrier. The following year, the U.S. Congress went so far as to create SEMATECH, a public-private consortium.

The outcome diverged a bit from the original intent. As the agreement drove up the prices of Japanese DRAM, an opening appeared in the market. Stepping into that opening was not the American companies that had pushed for the regulations, but South Korea, which was just taking its first baby steps.

On February 8, 1983, Samsung Chairman Lee Byung-chull announced the company's entry into the semiconductor business from a hotel in Tokyo. This was the so-called Tokyo Declaration. On November 7 of that year, Samsung succeeded in developing 64K DRAM. It took just nine months from the declaration, making Korea the world's third to do so, following the United States and Japan. Of course, Japan was already dominating the global market with that very product at the time.

The story of Samsung expanding facilities even during periods of price crashes is frequently cited in this industry. It meant riding the silicon cycle in reverse: building factories when prices hit rock bottom so that they were already up and running a few years later when prices rebounded. It was a gamble where a single miscalculation could drag down the entire company. In fact, many companies that took the same gamble later collapsed. What is indisputable is the result. In September 1992, Samsung became the first in the world to develop 64M DRAM, and that year, with a 13.5% market share, it surpassed Toshiba (12.8%) to claim the world's No. 1 spot in DRAM. In 1993, it became No. 1 across all memory semiconductors.

What came next was a game of chicken. It was a battle where, even as demand plunged, no one cut production—instead ramping it up and waiting for rivals with higher costs to collapse first. It sounds bizarre, but the calculation was clear. If you keep production lines running at prices where every unit sold is a loss, the lower-cost producer bleeds less while the higher-cost producer bleeds more. Hold out for just a few quarters, and a competitor vanishes, leaving their entire market share to the survivor.

That is exactly what happened. In 2007, DRAM prices plummeted 85%, and fell another 58% in 2008. In January 2009, Germany's Qimonda went bankrupt. It was the 5th-ranked company holding about 10% market share, and though the German government provided $500 million in aid, cumulative losses exceeded $3 billion. About 12,000 people lost their jobs.

Korea was not unscathed, either. In 1999, following the Asian financial crisis, Hyundai Electronics acquired LG Semicon for about $2.1 billion through a government-led "big deal," saddling the new company with nearly $10 billion in debt. In 2001, it changed its name to Hynix Semiconductor, and in February 2012, it became SK Hynix after SK Telecom acquired a 21% stake for 3.4 trillion won.

Elpida, the last bastion of Japanese DRAM, could not hold out. In February 2012, it filed for bankruptcy with 448 billion yen, about $5.5 billion, in debt. In 2013, Micron acquired the company, absorbing the remnants of Japanese DRAM.

In this way, the DRAM market effectively became a three-way split: South Korea's Samsung Electronics and SK Hynix, and America's Micron. While a few other players remained, such as Taiwan's Nanya, their share was modest. An industry that once boasted dozens of players had shrunk to a handful in little more than a decade. This oligopoly sets the stage for the next story.

Worth remembering The regulations struck not the intended target, but the bystander standing next to it—marking the moment leadership changed hands for the second time.

06How a Niche Graphics Card Technology Became the Center of the Industry

Memory competition had long been a battle over capacity. Cramming more cells into the same surface area was the very definition of skill. But at a certain point, the bottleneck shifted elsewhere. No matter how fast a computing unit ran, if the pathway bringing data in was too narrow, the processors sat idle. It was just like having an enormous warehouse with only a single door—you simply couldn't move goods in and out fast enough. The issue was no longer capacity, but bandwidth.

The solution that emerged was HBM (High Bandwidth Memory). In October 2013, JEDEC adopted a standard jointly developed by SK Hynix and AMD. The approach worked like this: instead of laying out individual chip slices (dies) sliced from silicon side by side across a substrate, they stacked them vertically, connecting the layers through microscopic wiring drilled directly through the silicon. That opened up a massive data highway of 1,024 bits per stack. On top of that, mounting this stack right next to the computing processor dramatically shortened the physical distance signals had to travel.

It is easy to misunderstand, but HBM is not built on an entirely new operating principle. The contents inside are still conventional DRAM; what changed was the packaging. That packaging, however, was notoriously difficult. Wafers had to be shaved thinner than paper, pierced with microscopic holes, and stacked atop one another with flawless alignment. If even a single layer was flawed, the entire stack had to be thrown out.

In the first quarter of 2015, SK Hynix kicked off mass production of HBM1, and it made its world debut inside an AMD graphics card. A single stack delivered 128 GB/s. Considering that GDDR5, commonly used for graphics cards at the time, ran at around 48 GB/s, the performance gap was undeniable. Even so, the industry's reaction remained lukewarm. Because it was so tricky to build and came with such a steep price tag, it was largely dismissed as a niche luxury meant only for a handful of high-end graphics cards.

That lukewarm reception dragged on for nearly a decade. In the meantime, the Great Memory Slump hit in 2023. Samsung Electronics' semiconductor division logged an annual operating loss of 14.88 trillion won, while SK Hynix posted an operating loss of 7.73 trillion won. It was a brutal year when chipmakers idled lines and slashed investments. With losses piling up in the trillions of won, keeping a line running for a product without a guaranteed customer base looked like pure recklessness to outside observers. It wasn't that those who pulled back were wrong; at the time, nobody could possibly know the answer. Yet even while bleeding cash, SK Hynix never let go of its HBM lines.

Meanwhile, the world outside was shifting. As deep-learning models swelled in size, the volume of values a processor had to read for every single calculation exploded. It reached a point where overall performance was determined not by raw computational power, but by how fast data could be shuttled back and forth. A bottleneck that had only troubled graphics cards ten years earlier was now dragging down the most expensive semiconductors on the planet.

The verdict on that gamble arrived swiftly. In 2024, SK Hynix secured a virtual monopoly supplying HBM3 for Nvidia's AI accelerators, and in March, it began mass production of the next-generation HBM3E. Then, in the first quarter of 2025, according to market research firm data at the time, SK Hynix captured a 36.9% share of DRAM revenue, overtaking Samsung Electronics at 34.4%. It was the first time since 1992 that the number one crown in DRAM had changed hands.

There were clear reasons for the dethroned leader, too. The formula that had kept Samsung on top for 33 years was churning out standardized DRAM cheaper and in higher volumes than anyone else on Earth. Its factories and its entire organization were finely tuned to stamp out identical commodities at immense scale. HBM, however, was an entirely different beast. Every customer demanded bespoke specifications, and stacking paper-thin dies resulted in notoriously low yields. To an organization built from the ground up to mass-produce standardized parts, HBM was an awkward, hesitant fit. Just as Japan could not let go of its ultra-high-reliability processes, this time Samsung found it difficult to abandon the playbook of standardized mass production that had made it number one.

What Samsung had done in the previous chapter, SK Hynix did this time around. A single production line kept alive when prices were at rock bottom shook a hierarchy that had held firm for 33 years.

Worth remembering The third turning point—this time between companies rather than nations, the exact same rule plays out once again.

07The Year Memory Eyes Half of Semiconductor Revenue

In the first quarter of 2026, global DRAM industry revenue reached $97 billion. That was an 81% surge from the previous quarter. In that same quarter, contract prices for conventional DRAM jumped 93–98% in just three months. All of this is based on TrendForce figures, and according to the same tally, DRAM revenue market share stood at 38.5% for Samsung Electronics, 28.8% for SK Hynix, and 22.4% for Micron.

Different research firms report different numbers. According to Counterpoint Research, DRAM market share in the second quarter of 2026 was 38% for Samsung Electronics, 25% for SK Hynix, 24% for Micron, 10% for CXMT, and 2% for Nanya. Looking at HBM alone, SK Hynix held 50%, Samsung Electronics 33%, and Micron 18%. In data from the same firm, NAND flash revenue in the first quarter of 2026 was $46 billion—about 3.5 times the figure from the same period a year earlier—with server SSDs accounting for 43% of that total. If you talk about rankings without specifying which firm and which quarter the numbers come from, you quickly end up with an inaccurate statement.

What is even more noteworthy than the numbers themselves is the speed at which rankings shift. In the first quarter of 2025, SK Hynix was ahead, but heading into 2026, both TrendForce's first-quarter tally and Counterpoint Research's second-quarter tally placed Samsung Electronics back at number one. It is a market where top positions can change hands in just a quarter or two. Because three companies divide the market among themselves, a single factory or a single new product from any one of them can immediately shake up the leaderboard.

There is a structural reason why prices surged like this. Producing the same capacity of HBM requires far more silicon wafer area than conventional DDR5. As Micron revealed in August 2026, it takes about 3 times as much silicon, and that gap widens with each new generation. Ramping up HBM production therefore directly reduces the manufacturing capacity available for standard DRAM. This is the exact channel through which headlines about AI servers translate directly into the price of RAM in your own computer.

The center of gravity has shifted as well. According to Gartner projections, global memory revenue in 2026 will reach $837 billion, with memory's share of total semiconductor revenue jumping from 27% in 2025 to 54% in 2026. To be sure, this is only the forecast of a single firm, and other market trackers at the time projected lower figures. Yet the direction is unmistakable. A product once dismissed as a "cheap commodity" has risen to claim half of the entire industry.

A fourth player has already stepped onto the board. China embarked on domestic memory development by founding CXMT in Hefei in May 2016 and YMTC in Wuhan in July of the same year. The United States allocated $52.7 billion for domestic manufacturing under the CHIPS Act in August 2022, and on October 7 of that same year, rolled out export controls against China that effectively blocked production facilities for DRAM below an 18 nm half-pitch and 3D NAND with 128 layers or more.

Yet those controls only blocked cutting-edge nodes; they could not prevent market entry altogether. CXMT went public on the Shanghai Stock Exchange on July 27, 2026, soaring 466% on its first day and raising 57.92 billion yuan. A company that started from scratch had become the fourth-largest DRAM maker and a darling of the capital markets in just ten years.

Worth remembering The real story behind today's headlines on RAM prices and market share—and the fact that a fourth player has already entered the game.

08What to Watch When the Next Round Begins

You don't need to try and predict the future. Instead, guided by the rules this story has shown us, let's look at what signs to keep an eye on to spot the shifting tide before anyone else.

First is the generational shift in HBM. HBM4 aims to widen the bus to 2,048 bits, delivering about 2 TB/s per stack. In this industry, the rankings have always been overturned at the turning points between generations. Whenever a new standard arrives, the production experience built up until then is reset all at once. Whichever maker can stably produce the new generation first takes the lion's share for the next few years.

Second is the production capacity consumed by HBM. The more wafers AI consumes, the less memory is left for PCs and smartphones. The direction of prices will depend on how much of this strain the market can endure, or whether chipmakers expand capacity enough to satisfy both sides.

Third is China. When output increases from mature nodes outside export controls, the cheapest tier wavers first. You just have to watch whether prices collapse in the lower-end product segments first. Rather than a race of technology, this is a contest of how long you can endure losses. It is the very same battle this industry has fought time and again.

Fourth is the next downturn. Samsung Electronics and SK Hynix have warned that the AI-driven supply shortage could last at least through 2027, and possibly until around 2030, while customers are already securing volumes for 2027 in advance. Yet the silicon cycle has never once disappeared. Expansion decisions made today will only emerge as finished goods years down the line, and no one can guarantee that demand will look like this when that time comes. When prices take a dive, who will keep on investing? Up until now, the next champion has always been decided right there. Whether that holds true this time around will be well worth watching.

And one final point. Most of those pushed out of this story were not cast aside because of poor manufacturing skills. Intel was the company that turned DRAM into a commercial product, Toshiba was the company that invented NAND flash, and Japan led the world in quality above all others. Even so, when the landscape shifted, they could not let go of the very formulas that had brought them victory. No one knows yet whose turn it will be next.

Worth remembering This is not a forecast, but a spectator's guide: generational shifts, capacity allocation, China, and the next downturn—these are the four things to watch.