Number Research
← Back to Explore
Economics Business History Corporate Reporting and Surveillance

From Sugar and Textiles to Silicon Supremacy: How Samsung Conquered DRAM and HBM

The contrarian 1983 bet that turned a watch-chip assembler into the king of DRAM and HBM

S.J. Nam 14 min read
From Sugar and Textiles to Silicon Supremacy: How Samsung Conquered DRAM and HBM

When people today think of Samsung, they picture smartphones and televisions. But the true engine of the company's global power sits invisibly inside almost every electronic device on Earth: memory chips. Samsung Electronics is the world's largest manufacturer of DRAM, the volatile memory that lets computers, phones, and data centers hold data while they work. That dominance was not inevitable. It was the product of a contrarian bet made in the early 1980s—one that seasoned observers, including many inside Samsung itself, considered reckless.

And yet the story does not end with victory. Four decades after that bet, the AI revolution nearly caught Samsung flat-footed. A specialized product called high-bandwidth memory (HBM) briefly turned the industry's perpetual champion into a chaser—and forced the company to relive, in compressed form, the same drama of humiliation, mobilization, and comeback that defined its origins. This is the full story.

A Late and Unlikely Entrant

South Korea in the 1970s was an improbable birthplace for a semiconductor titan. The country had emerged from the Korean War as one of the poorest nations on the planet, with a per-capita income lower than much of sub-Saharan Africa. Heavy industry—steel, shipbuilding, chemicals—defined its export ambitions under President Park Chung-hee's development drive. Semiconductors, an industry where a single fabrication plant could cost more than the annual budget of a Korean ministry, seemed hopelessly out of reach.

Samsung's first real toehold in silicon came in 1974, when the group acquired a controlling stake in Korea Semiconductor Co., a small, struggling firm with the capacity to make simple integrated circuits for watches and calculators. The purchase was financed in part by Lee Kun-hee, the son of founder Lee Byung-chul, reportedly with his own money—much of the group's leadership was skeptical that a company built on sugar refining, textiles, and consumer electronics assembly had any business in chips. For nearly a decade, that operation produced only basic components. Samsung had a factory and some know-how, but it was nowhere near the frontier of memory technology, which was then owned by American giants like Intel (which had invented the DRAM) and a rising cohort of Japanese firms such as NEC, Hitachi, Fujitsu, and Toshiba.

The Tokyo Declaration

The decisive turning point came on February 8, 1983. Lee Byung-chul, by then in his seventies and running one of Korea's largest business empires, issued what became known inside the company as the "Tokyo Declaration." From Tokyo, he announced that Samsung would commit itself fully to the DRAM business and mass-produce advanced very-large-scale integrated circuits.

The decision was audacious to the point of seeming irrational. DRAM was a brutally cyclical commodity business requiring enormous, continuous capital investment in fabrication plants, where each new chip generation demanded fresh billions of dollars and ever-finer manufacturing precision. Samsung had no leading-edge process technology, no established customer base in memory, and no track record at the frontier. The Japanese firms that dominated the market were widely regarded as unbeatable, having just displaced American producers through relentless quality and scale—Intel itself would exit the DRAM business in 1985, conceding defeat. Critics at home and abroad argued Samsung would burn cash it could not afford to lose; some Japanese executives openly dismissed the Korean effort as fantasy.

Lee's reasoning was strategic rather than sentimental. He judged that memory chips would become the foundational raw material of the coming information age—a product with effectively limitless demand—and that Korea's disciplined, well-educated workforce and the government's willingness to back "national champion" industries gave Samsung a fighting chance. He also understood that the alternative, remaining a low-margin assembler of others' designs, was a dead end. Resource-poor Korea, he argued, had to sell brainpower embedded in silicon.

Cracking the 64K DRAM in Record Time

Having declared its intent, Samsung had to actually build a chip. The company licensed the 64K DRAM design from Micron Technology—then a small Idaho startup—and process technology from Zytrex, and dispatched engineers to the United States and Japan to absorb everything they could, often under difficult and secretive conditions, since incumbents had no interest in training a future rival. Korean-American engineers recruited from Silicon Valley became a critical conduit of tacit knowledge. Teams famously worked punishing schedules, and company lore celebrates development engineers who marched between facilities overnight as a ritual of resolve.

The results stunned the industry. In November 1983, only months after the Tokyo Declaration, Samsung succeeded in developing a working 64K DRAM—becoming the third country-team in the world, after the United States and Japan, capable of the feat. A technology gap that experts measured in ten or more years had been compressed to roughly four. Samsung simultaneously threw up its first dedicated fabrication plant in Giheung (Kiheung), south of Seoul, in about six months—construction of that kind ordinarily took eighteen.

The chip itself was already a generation or two behind the market leaders when it arrived, and Samsung initially lost money on nearly every unit it sold. But the achievement proved the organization could learn and execute at extraordinary speed—the single most important capability in an industry defined by relentless generational transitions.

Surviving the Downturn and Catching Up

The mid-1980s tested Samsung's resolve almost immediately. A severe global glut—worsened by aggressive Japanese capacity expansion—sent DRAM prices crashing far below production costs. The price of a 64K DRAM collapsed from several dollars to well under a dollar, and Samsung's semiconductor division hemorrhaged money for years; by some accounts the accumulated losses through 1986 nearly equaled the company's entire paid-in capital. Many companies would have retreated. Intel and most American producers did exactly that.

Instead, Samsung did the opposite of conventional wisdom: it kept investing through the downturn, pouring capital into new lines and next-generation designs precisely when rivals were pulling back. This counter-cyclical strategy—later a hallmark of the entire Korean memory industry—meant Samsung emerged from each slump with newer plants and lower costs than competitors who had hesitated. Fortune smiled, too: the 1986 U.S.–Japan Semiconductor Agreement, which forced Japanese firms to raise prices and restrain exports, opened a window in the American market that Samsung drove straight through. When DRAM prices rebounded sharply in 1987–88, the division swung from ruinous losses to enormous profits almost overnight.

The company moved steadily up the technology ladder, mastering the 256K generation and then the 1-megabit DRAM in the second half of the decade, each time narrowing the Japanese lead. Government-backed national R&D consortia, cheap capital circulating within the chaebol structure, and a willingness to endure years of losses in pursuit of scale all reinforced the push. In 1992, Samsung developed the world's first 64-megabit DRAM—no longer a follower but the pacesetter.

Becoming the DRAM Giant

By the early 1990s, the persistence paid off spectacularly. Samsung began matching, then beating, its Japanese rivals to the newest chip generations, and in 1992 it rose to become the world's single largest producer of DRAM. It has held the top position essentially ever since—a run of dominance now stretching more than three decades, through the PC era, the mobile revolution, and the cloud buildout.

The industry structure Samsung helped create endures today. Alongside its Korean compatriot SK Hynix, Samsung anchors an industry in which two South Korean firms manufacture roughly two-thirds of the world's memory chips. The Japanese firms that once looked invincible largely exited or consolidated—NEC and Hitachi merged their DRAM units into Elpida, which went bankrupt in 2012 and was absorbed by Micron—unable to match the Korean model of relentless, deep-pocketed, counter-cyclical investment. Lee Kun-hee, who succeeded his father, drove the culture further with his 1993 "New Management" declaration in Frankfurt—"change everything except your wife and children"—cementing an obsession with quality and generational leadership.

For thirty years, the formula seemed unbeatable. Then the definition of "memory leadership" quietly changed.

The HBM Blind Spot: How the Champion Missed the AI Wave

High-bandwidth memory was born not in Suwon but in Icheon, at SK Hynix. Working with AMD, SK Hynix developed the world's first HBM in 2013: instead of spreading DRAM chips flat across a circuit board, HBM stacks them vertically—connected by thousands of microscopic vertical channels called through-silicon vias (TSVs)—and places the stack right beside the processor. The result is vastly wider data bandwidth at lower power, exactly what a GPU crunching enormous parallel workloads needs.

For years, however, HBM looked like a niche curiosity. It was expensive, hard to manufacture, and served a small market of graphics cards and supercomputers. Samsung participated—it mass-produced HBM2 in 2016 and for a time was a leading supplier—but the product never seemed worth the trouble for a company printing money on commodity DRAM. Around 2019, with the market stagnant, Samsung reportedly scaled back its dedicated HBM development effort, judging that the payoff didn't justify the resources. It was a defensible spreadsheet decision. It was also, in hindsight, the most expensive miscalculation in the company's semiconductor history.

SK Hynix made the opposite bet. It kept its HBM teams intact, refined its proprietary MR-MUF packaging process, and worked closely with Nvidia. When ChatGPT detonated the generative-AI boom in late 2022, every Nvidia H100 accelerator needed stacks of the most advanced HBM available—and SK Hynix was effectively the only qualified supplier of HBM3. The economics inverted overnight: HBM commanded several times the price of conventional DRAM per bit, and the perennial number-two in memory suddenly owned the most profitable product in the industry. By 2024–25, SK Hynix's operating profits were rivaling and at times exceeding Samsung's entire semiconductor division—something previously unthinkable.

The Painful Years: Failed Qualifications and a Leadership Shake-Up

What followed was the most public humiliation in Samsung Memory's modern history. Samsung raced to qualify its fifth-generation HBM3E chips with Nvidia, the gatekeeper customer of the AI era—and repeatedly failed. Through 2024, reports emerged of test setbacks tied to heat and power-consumption issues in Samsung's 8-layer and 12-layer HBM3E stacks. Each rumored failure knocked billions off Samsung's market value; each report of SK Hynix's sold-out order books added to the sting. Micron, long the distant third player, also leapfrogged Samsung into Nvidia's HBM3E supply chain.

The crisis triggered soul-searching about deeper problems: an engineering culture that had grown bureaucratic, a talent drain to rivals, and a legacy of prioritizing cost-down commodity manufacturing over bleeding-edge customization. In May 2024, Samsung abruptly replaced its semiconductor chief, installing memory veteran Jun Young-hyun to lead the division. In an unusually blunt message to employees, leadership acknowledged the company was in crisis and apologized to shareholders for disappointing results. Later reorganizations consolidated scattered engineering resources, placing HBM development directly under a unified DRAM design organization led by Hwang Sang-jun, the company's high-value DRAM specialist.

Samsung was, in effect, running its 1983 playbook against itself: the company that had once been the hungry, disrespected challenger now had to remember how to be one.

The Comeback: HBM3E Breakthrough and the HBM4 Counterattack

The turnaround came in stages. In September 2025, after roughly eighteen months of failed attempts, Samsung finally passed Nvidia's qualification tests for its 12-layer HBM3E, and shipments to Nvidia began in the third quarter of 2025. The company confirmed that HBM3E was in mass production and selling to all major customers, and its memory business posted record quarterly revenue of 26.7 trillion won on the back of resurgent AI-server demand. The symbolism deepened in late October 2025, when Nvidia CEO Jensen Huang met Samsung Executive Chairman Jay Y. Lee in Seoul during the APEC summit—famously sharing fried chicken and beer—and the two companies announced plans for an "AI megafactory" in which Samsung would deploy some 50,000 Nvidia GPUs to run AI across its own chip design and manufacturing operations.

But HBM3E was a battle already largely lost to SK Hynix. The real counterattack was HBM4, the sixth-generation product designed for Nvidia's next flagship platform, Vera Rubin. Here Samsung made an aggressive architectural gamble that echoed the old Tokyo Declaration spirit: it built HBM4 on its most advanced sixth-generation 10-nanometer-class (1c) DRAM process—a full node ahead of competitors—and manufactured the logic base die on its own 4-nanometer foundry line, freeing it from dependence on TSMC. It also leaned on its unique position as the only company in the world combining memory, logic design, foundry, and advanced packaging under one roof.

The gamble paid off. Samsung delivered HBM4 samples to Nvidia in September 2025, entered the final qualification phase, and by December 2025 Korean media reported its chips had scored the highest marks among all memory makers on speed and power efficiency. In January 2026, reports indicated Samsung had cleared final qualification with both Nvidia and AMD—notably without requiring a redesign even after customers raised their performance targets—and would begin shipments in February 2026. Its HBM4 runs at 11.7 gigabits per second per pin, comfortably above the roughly 10 Gbps that Nvidia demanded and well beyond the 8 Gbps JEDEC baseline, with headroom toward 13 Gbps. Preorders for 2026 HBM supply reportedly exceeded Samsung's planned output before the year even began, and the company laid out one of its largest DRAM expansions ever, targeting up to 150,000 wafers per month of 1c capacity for HBM4 by the end of 2026.

Where Things Stand Today (Mid-2026)

At Nvidia's GTC 2026 conference in San Jose this March, Samsung's exhibit told the story of a company that believes it has re-seized the initiative. The centerpiece was HBM4—now in mass production and designed into the Vera Rubin platform—alongside the first public showing of its seventh-generation HBM4E, targeting 16 Gbps per pin and 4.0 terabytes per second of bandwidth per stack. Samsung also previewed hybrid copper bonding, a next-generation stacking technique intended to enable 16-high (and taller) HBM stacks while cutting thermal resistance by more than 20 percent—critical as Nvidia pushes suppliers toward 16-layer HBM as early as late 2026.

The competitive picture remains fierce. SK Hynix still holds the largest share of the HBM market and completed its own HBM4 development first, in September 2025; Micron is pressing from behind; and each Rubin-class accelerator consumes eight HBM4 stacks, making supply commitments and yields—Samsung's 1c yields were still maturing through the ramp—the decisive battleground. Per-stack prices in the mid-$500 range, more than 50 percent above HBM3E, mean the stakes are measured in tens of billions of dollars annually. The next frontier is custom HBM, in which memory makers co-design base logic dies with individual customers like Nvidia, Google, and Broadcom—a shift that plays directly to Samsung's integrated memory-plus-foundry model.

Conclusion

Samsung's ascent from a small watch-chip assembler to the king of DRAM is less a story of a single breakthrough than of institutional willpower. The 1974 acquisition gave it a doorway; the 1983 Tokyo Declaration gave it a mission; the record-breaking 64K DRAM proved it could compete; and years of stubborn investment through crushing losses turned a long-shot bet into an unassailable lead.

The HBM saga is the same story refracted through a new era—with the roles briefly reversed. The champion grew comfortable, dismissed a "niche" technology, and watched a hungrier rival seize the most valuable memory market in history. What is striking is how Samsung responded: not with denial, but with the old formula—leadership upheaval, organizational mobilization, a technically audacious leap (1c DRAM plus an in-house logic die), and massive counter-cyclical capacity investment ahead of certain demand. Whether Samsung ultimately dethrones SK Hynix in HBM remains an open question in 2026. But the arc from the Tokyo Declaration to the HBM4 comeback carries one consistent lesson: in memory, no lead is permanent, and the company willing to bet biggest and endure longest tends to write the next chapter.

References

Related Essays

The $340 Demo That Became a $61,000 Bill
Economics • Economics of Information Systems

The $340 Demo That Became a $61,000 Bill

Everyone watches the token price. It's the smallest, most-discounted, least-honest number in your stack — and it explains almost nothing about what an AI system actually costs to run.

S.J. Nam, Ph.D Jul 5, 2026 7 min read
Immigration, Innovation, and Wages in the U.S.
Economics • Microeconomics

Immigration, Innovation, and Wages in the U.S.

In this article, we’ll explore how immigration affects both innovation and wages in the U.S., drawing on decades of studies. We’ll also compare how different researchers have tried to measure these effects—sometimes reaching very different conclusions. The goal is to make complex economics research accessible, so you can see what the evidence shows and why scholars sometimes disagree.

S.J. Nam, Ph.D Jun 13, 2026 11 min read