Intel's 40-Year Quest to Reclaim the Memory Market

A forgotten rivalry wakes up

In 1985, Intel abruptly exited the DRAM business. The company had pioneered the technology in the early 1970s, producing the original 1103 chip — the first commercial dynamic random-access memory device — at a time when memory was treated as the central nervous system of computing. But Japanese competition, particularly from NEC and Toshiba, drove prices into the ground through relentless scale and state-supported investment. Intel redirected its resources toward the microprocessor, a bet that defined the next four decades of computing and made the company the most valuable semiconductor firm on Earth.

Now, under CEO Lip-Bu Tan, Intel is circling back — not with the blunt industrial aggression of the 1980s, but with a surgical approach aimed at the bleeding edge of AI infrastructure.

Speaking on Michael Marks’ TechSurge podcast, Tan suggested the company is exploring new ways to integrate memory directly onto CPU packages, framing it as an extension of work already underway rather than a full-scale re-entry into commodity DRAM. He pointed to Intel’s Lunar Lake platform, which already places memory within the same package as the processor using a custom SoC architecture, as evidence that stacked DRAM is on the company’s roadmap. The implication, carefully left unstated, is that Intel may be preparing for a second act in the memory business it essentially created.

The 1985 exit that shaped an industry

To understand the significance of Intel’s current maneuvering, it helps to remember what leaving the memory business cost the company — and what it gained. In the early 1980s, Intel’s memory division was its crown jewel. The 1103 DRAM chip powered mainframes and minicomputers across the emerging personal computer industry. By 1984, Japanese manufacturers were flooding the market with cheaper, equally reliable memory chips, backed by subsidies from the Japanese Ministry of International Trade and Industry. Intel’s gross margins on DRAM had collapsed from above 50 percent to single digits.

Then-CEO Andy Grove made one of the most consequential decisions in corporate history: he sold off the memory business and focused exclusively on the microprocessor. The logic was stark — Intel could compete on innovation in processors, but it could not win a price war against subsidized Japanese manufacturers. Within a decade, that decision had made Intel the dominant force in personal computing. But it also created a vacuum in the memory space that would eventually be filled almost entirely by Korean firms.

Why now? Because AI broke the old rules

The memory market of the 1980s was a textbook commodity trap — prices swung with cycles of overproduction and demand, margins collapsed, and the winner was whoever could manufacture cheapest. Intel lost that race to Japan and walked away. The economic logic was merciless: memory is memory, and the market pays for performance per dollar, not brand loyalty or architectural elegance.

The market looks radically different today, and not simply because AI exists. AI workloads have transformed memory from a generic component into a critical bottleneck. Traditional DRAM scaling — making cells smaller, denser, cheaper — has hit physical and economic walls. The solution that emerged is not cheaper memory but smarter memory placement: stacking DRAM layers directly on or near the processor, reducing the distance data must travel, and dramatically increasing bandwidth.

High-bandwidth memory (HBM), developed primarily by SK Hynix and Samsung, represents the leading edge of this shift. HBM uses through-silicon vias to stack multiple DRAM dies vertically, achieving bandwidth that conventional DIMM-based memory cannot approach. It is used in NVIDIA’s GPU accelerators, AMD’s MI-series AI chips, and increasingly in custom silicon from Google, Amazon, and Microsoft. HBM commands premium pricing — often three to five times the per-gigabyte cost of standard DRAM — and tight supply. Demand continues to outstrip capacity, with industry observers projecting persistent shortages through at least 2030. TSMC’s acquisition of SkyWater and its push into advanced packaging further validated that the industry’s center of gravity is shifting toward integrated memory solutions.

Tan’s comments reflect this structural shift. He stated plainly that current memory architecture still has enormous room for innovation — a direct challenge to the status quo dominated by two Korean firms and their proprietary HBM stacks. The framing matters: Intel is not talking about competing on commodity DRAM price. It is talking about redefining where memory sits in the system hierarchy.

The SK Hynix connection

The most concrete signal came not from Tan’s words but from his hiring decisions — the kind of move that reveals strategic intent more reliably than any press release. Intel recruited Seok-Hee Lee as executive vice president overseeing foundry operations and advanced packaging technology. Lee is far from a typical executive hire in the semiconductor industry.

He spent 2000 to 2010 at Intel in various engineering and operations roles before moving to SK Hynix, where he served as CEO from 2016 to 2020. During his tenure at SK Hynix, Lee played a central role in developing and scaling HBM — the memory technology that powers the world’s most powerful AI chips. His expertise sits at the exact intersection of advanced packaging and high-performance memory that Intel appears to be targeting. More importantly, he brings relationships with Korean supply chain partners that Intel has spent years trying to build from scratch.

This is not coincidental. Intel’s foundry business needs credible technology leadership, and hiring someone who has successfully managed the production of HBM — arguably the most technically demanding memory product on the market — sends a signal to both customers and competitors. Lee brings the design pedigree and the manufacturing relationships that could accelerate Intel’s push into memory-adjacent products. His presence also suggests that Intel’s approach will lean heavily on advanced packaging rather than attempting to replicate Samsung or SK Hynix’s vertical DRAM fabs — a more capital-efficient path that plays to Intel’s strengths.

Who wins, who loses

The Korean duopoly — Samsung and SK Hynix — controls roughly 90 percent of the global DRAM market and an even larger share of the HBM segment. Their dominance is built on scale, process maturity, and deep customer relationships with every major server and PC manufacturer. Samsung invests more in memory R&D than most countries spend on defense. SK Hynix maintains an almost exclusive relationship with NVIDIA, supplying the HBM that powers the majority of AI training clusters worldwide.

Intel entering even a niche segment of memory production would apply pressure on pricing and force competitors to defend territory they have held unchallenged for decades. Samsung, already grappling with a brutal commodity cycle and massive capital expenditure requirements for its own advanced memory projects — including its GAA (gate-all-around) transistor process targeted at next-generation DRAM — would face a new rival with deep pockets and a brand that still carries significant weight in the data center. Samsung’s memory division has been cyclical suffering; an Intel entry could accelerate margin compression just as the company is trying to fund its most ambitious technology transitions.

SK Hynix would feel the threat differently and perhaps more acutely. Its relationship with NVIDIA is both its greatest asset and its primary vulnerability. If Intel’s stacked memory approach — integrating DRAM directly onto the processor package — reduces the need for discrete HBM in certain workloads, it could undermine SK Hynix’s most profitable customer relationship over time. NVIDIA itself has explored alternatives to SK Hynix-supplied HBM, including working with Samsung and developing its own memory interconnect standards. Intel’s entry adds another variable to an equation that is already shifting.

Second-order effects: supply chain realignment

The implications extend well beyond market share. A third credible supplier in advanced memory would reshape procurement strategies for cloud providers and AI chip designers who currently have only two meaningful options. Currently, companies building training clusters negotiate HBM allocation with Samsung and SK Hynix from a position of weakness — lead times stretch months ahead, and pricing is set by sellers. Even a late entrant like Intel could improve leverage by expanding the available supply base and introducing competitive tension.

This realignment would also affect equipment suppliers. ASML, Tokyo Electron, and Lam Research have built their memory-related revenue around Samsung and SK Hynix’s fab expansion plans. An Intel memory strategy would redirect some of that capital expenditure toward Intel’s own facilities, potentially in the United States or Europe — aligning with the CHIPS Act and similar industrial policy initiatives. The geopolitical dimension matters: the current memory concentration in South Korea represents a strategic vulnerability that U.S. and EU policymakers have been actively trying to address.

There are also implications for Intel itself. Committing to memory production would require significant foundry investment at a time when the company is already running foundry operations at a loss. But it could also create a new revenue stream that justifies the foundry bet and gives Intel a foothold in the AI infrastructure supply chain that goes beyond just manufacturing other companies’ chips.

The unanswered questions

Tan’s statements were deliberately vague. He did not confirm a formal DRAM product line, nor did he specify timelines or technical specifications. There is also ambiguity about whether the initiatives he described are corporate programs tied to Intel’s foundry business or connected to Walden International, the venture capital firm he founded and still chairs. Walden has made significant investments in memory and advanced packaging startups in recent years, including stakes in companies working on novel memory architectures and chiplet-based integration. The overlap between Tan’s VC portfolio and his Intel strategy raises questions — however fair or unfair — about whether Intel’s memory ambitions are being shaped by broader ecosystem investments.

Intel’s foundry business, while growing and securing some high-profile customers, remains unproven at the cutting edge. Competing with Samsung and SK Hynix on memory manufacturing would require billions in investment and years of process development — resources Intel has struggled to allocate consistently under previous leadership. The company’s idm 2.0 strategy promised foundry success, but financial results have been uneven, and the memory business operates on thinner margins and longer cycles than most other semiconductor segments.

There is also the question of timing. Memory development cycles are measured in years, not quarters. Even if Intel commits seriously today, meaningful product shipments would likely not arrive until 2027 or later — by which time the HBM market may have consolidated further around Samsung and SK Hynix’s next-generation offerings. The window for disruption is narrow.

What happens next

If Intel commits seriously to memory, the impact will be felt first in the AI infrastructure supply chain, where companies are already wrestling with HBM allocation, lead times, and single-source dependencies. A third major supplier — even a late one — could improve leverage for cloud providers and chip designers who have had only two meaningful options for years. It could also accelerate the industry’s migration toward chiplet-based architectures, where memory is treated as a modular component rather than a monolithic stack.

The timeline matters. Tan’s hints suggest anything concrete is years away. But the mere possibility is already reshaping conversations in boardrooms from Seoul to Santa Clara. Samsung and SK Hynix are monitoring Intel’s foundry progress closely. NVIDIA is evaluating whether to diversify its HBM supply chain further. Cloud providers are quietly exploring alternative memory architectures that reduce dependence on any single supplier.

Intel left the memory business 40 years ago because it could not win a price war against subsidized Japanese manufacturers in a commodity market. This time, the war is not about price — it is about bandwidth, latency, and physical proximity. The economics have changed. The technology landscape has changed. The geopolitical context has changed. Whether Intel can execute with the discipline and investment that its 1985 exit demonstrated it lacked remains the question that will determine whether this comeback story has a second act or becomes another chapter in Intel’s long struggle to translate ambition into execution.