Samsung's Quiet DRAM Strategy Shift Could Redraw AI Chip Supply

Samsung Is Playing a Different Game

Samsung Electronics is winning the wrong battle. The world’s largest DRAM maker is projected to grow its production capacity by just 6% this year and a mere 1.3% next year, according to data from Omdia obtained by Chosunbiz. Meanwhile, rival SK Hynix is sprinting ahead with roughly 10% annual growth, steadily eroding Samsung’s decades-old leadership margin.

The reason is not a lack of capital or demand. It is a deliberate choice. Samsung is prioritizing the transition from its 10-nanometer fifth-generation (1b) process to its sixth-generation (1c) process and stabilizing yields on those newer nodes over simply pumping out more wafers. The strategy, rooted in Vice Chairman Jeon Yong-hyeon’s long-standing philosophy, trades near-term volume for what the company believes is sustainable competitiveness in the era of next-generation AI chips.

The Numbers Tell a Starker Story

Samsung’s DRAM production capacity is expected to rise from 7.695 million wafers last year to 8.175 million this year — a 6% increase. Next year, it climbs only to 8.28 million wafers, barely above this year’s level. Those figures look thin against the backdrop of what the market had been pricing in: a semiconductor industry running at full throttle to serve AI data centers hungry for memory.

The gap with SK Hynix is where the real alarm sits. SK Hynix’s capacity is projected to grow from 6.06 million wafers in 2025 to 6.66 million in 2026 and 7.29 million in 2027. By 2027, Samsung’s lead over its rival will have shrunk from 163,500 wafers — a 27% advantage — to just 990,000 wafers, a 13.6% edge. That is a roughly 39% compression in two years.

For a company that has built its memory dominance on scale, that trajectory is uncomfortable.

Retrofit, Not Rebuild

Much of Samsung’s near-term output increase is coming from retrofitting existing lines at its Pyeongtaek campus rather than breaking ground on new fabs. The Pyeongtaek facility is expected to account for 55.4% of Samsung’s total DRAM capacity by 2027, up from 49.5% in 2025. But retrofitting is slower and riskier than building new capacity from scratch. Existing equipment must be modified to handle the 1c process, and that conversion process itself can generate production losses — wafers that go bad during the transition and never make it to shipment.

There is also a structural reshuffling happening at Samsung’s older sites. The facility in Hwaseong, which for years contributed a steady share of DRAM output, has seen stagnant or declining capacity for three consecutive years. Samsung already shut down two-dimensional NAND production at its Hwaseong Line 12 and converted that line’s back-end operations into an end-fab supporting Line 15’s DRAM output. Lines 16 and 17 are caught in the same overhaul cycle. Legacy DRAM and NAND used to account for 30% of Samsung’s memory revenue — a proportion the company has publicly committed to shrinking.

The net effect is that Samsung’s headline capacity growth underestimates both the pain of transition and the upside of denser chips. Newer process nodes pack more bits per wafer. So even if wafer input grows slowly, bit-level output can expand faster. An industry source noted that Samsung’s apparent slowdown looks worse in wafer terms than in actual bit supply. Still, wafer-level metrics are what markets watch, and the perception gap matters.

The HBM Question

The strategic logic behind Samsung’s restraint becomes clearer when you look at where the money is going. The 1c process is not just a DRAM upgrade. It is the process that will power sixth-generation high-bandwidth memory — HBM4 — the critical component inside the AI accelerators running everything from large language models to inference workloads. HBM is where margins are, where customers like Nvidia and Amazon are signing multi-year supply deals, and where SK Hynix currently holds a commanding lead.

Samsung’s decision to funnel capital and engineering attention into 1c and yield stabilization is a bet that it can catch up in HBM while avoiding the pitfalls of overexpansion. The company stumbled in yield competition on its fifth-generation DRAM process, a weakness that became painfully visible to customers. Repeating that mistake at the sixth generation would be costly in more than revenue — it would damage trust with the very buyers who matter most.

But the bet carries risk. While Samsung slows to stabilize, SK Hynix is adding capacity on multiple fronts and deepening its HBM relationships. Every quarter Samsung holds back on volume is a quarter its rival locks in another design win.

Who Wins, Who Loses

SK Hynix is the clear short-term beneficiary. Its double-digit capacity growth, combined with its existing HBM advantage, positions it to capture more of the surging AI memory demand this year and next. Customers facing supply constraints will find fewer reasons to split orders between the two Korean makers.

Samsung’s legacy DRAM business is the loser in this transition. The company has acknowledged that older process nodes are no longer profitable enough to justify their share of capacity. The shutdown of Lines 16 and 17’s traditional output is part of a planned retreat, not an accident.

The broader loss is reputational. Samsung has long marketed itself as the memory giant that can deliver at scale. A strategy that consciously accepts slower growth in exchange for better yields sends a mixed signal to buyers who need both. If the company’s 1c yield targets are not met on schedule — and no one outside Samsung knows whether they will be — Samsung could end up with neither the volume nor the reputation advantage it was trying to secure.

What Happens Next

Memory market analysts expect the DRAM supply shortfall to persist through the second half of the year and likely beyond, extending the current pricing cycle that has been extraordinarily profitable for the entire Korean semiconductor industry. That extended window is Samsung’s greatest opportunity: more time to perfect the 1c process, more time to qualify for HBM4 supply, more time to close the gap with SK Hynix without the pressure of immediate volume loss.

Whether that patience pays off depends on a single metric that will not be publicly available for months: Samsung’s 1c yield rate. If it stabilizes quickly, the strategy looks prescient. If it drags, Samsung’s capacity stagnation looks like hesitation at exactly the wrong moment.

The company’s own words suggest it is comfortable with that gamble. A Pyeongtaek campus representative told Chosunbiz that scaling production aggressively with currently operational lines is difficult and that the fundamental policy remains fixed on yield and productivity rather than quantity. That is not the language of a company worried about losing ground. It is the language of a company playing a longer game.

The question is whether the longer game is the right one.