In 2025, SK Hynix overtook Samsung in annual operating profit for the first time in either company’s history — 47.2 trillion won versus Samsung’s 43.6 trillion won. SK Hynix makes memory chips. Samsung makes everything from smartphones to foundry chips to home appliances. That a focused memory specialist outearned the conglomerate comes down to one product: High Bandwidth Memory.
The HBM market is a near-duopoly. Two South Korean companies control approximately 78–92% of all HBM supply. When one of those players holds roughly 70% share and has sold out all 2026 production, the ripple hits every infrastructure budget worldwide. This article maps how SK Hynix built that lead, why Samsung fell behind, and what it means for enterprise buyers planning hardware spend. It’s part of the AI memory crunch series.
How did the HBM duopoly form and why do just two companies control AI memory supply?
SK Hynix and Samsung together hold approximately 78–92% of all HBM revenue. Counterpoint Research‘s Q3 2025 data puts SK Hynix at 57% and Samsung at 22%. TrendForce estimates SK Hynix climbed to roughly 70% by Q1 2026. Micron holds the remaining 21–24% — meaningful, but structurally behind in the AI-grade qualification cycles that actually matter.
It wasn’t always this concentrated. In 2009, ten companies competed in DRAM. The 2011 downcycle triggered terminal consolidation: SK Telecom acquired Hynix in 2012, Micron purchased bankrupt Elpida in 2013. Within five years it was a triopoly.
Re-entry isn’t realistic. A greenfield DRAM fab costs $15–25 billion with a four-to-five year build time, plus 18–36 months of yield ramp. HBM packaging adds another $2–5 billion per node. Total cumulative capex to compete at meaningful scale runs $30–50 billion over seven to ten years. Chinese manufacturers CXMT and YMTC get cited as challengers regularly — they’re three to five nodes behind and blocked from EUV lithography by export controls. This structure holds for at least a decade.
IDC frames it directly: 2026 DRAM supply growth is expected at only 16% year-on-year versus 20–30% historical norms, because wafer capacity has been deliberately reallocated to higher-margin HBM. This isn’t a self-correcting shortage. It’s a permanent structural shift.
Why does SK Hynix lead in HBM? The TSV yield advantage explained
SK Hynix’s dominance comes down to a manufacturing process difference that compounds mathematically across every wafer.
HBM is built by stacking 12 DRAM dies vertically using Through-Silicon Vias — copper-filled channels drilled through each die, roughly 5–10 micrometres wide.
💡 A Through-Silicon Via (TSV) is a conductive channel drilled straight through a silicon chip, allowing electrical signals to travel vertically between stacked chips rather than across longer horizontal interconnects.
Here’s why that matters. An HBM3E 12-Hi stack contains 1,200–1,800 TSVs per layer. One defective TSV kills the entire stack. At 99% per-die yield, a 12-die stack yields 88.6% usable. Drop to 95% per-die yield and you get 54%. A 4-point improvement in per-die yield translates to a 34-point improvement in stack yield. Small differences at the die level determine who wins the market.
SK Hynix uses MR-MUF — Mass Reflow with Molded Underfill. All 12 dies bond and seal in a single thermal cycle, reducing voids and raising yield. Their 1bnm DRAM node runs at 75–80% stack yield.
Samsung uses TC-NCF — Thermal Compression with Non-Conductive Film — where adhesive film is applied layer by layer. Historically slower, historically lower yield. Their “Advanced TC-NCF” variant narrowed the gap but didn’t close it. Samsung’s 1cnm node delivers 60–65% stack yield.
That 10–15 point gap, compounded across 12 dies, is the manufacturing basis for SK Hynix’s market position: more usable stacks per wafer, lower defect rates, faster Nvidia qualification. For a full engineering treatment, the technical architecture that explains SK Hynix’s TSV advantage goes deeper.
What were Samsung’s HBM3E quality problems and how consequential were they?
Samsung’s HBM3E failure wasn’t a single misstep. It was a two-year compounding sequence.
It started with thermal failures during Nvidia qualification in 2023. Samsung’s TC-NCF stacking produced higher thermal resistance than Nvidia’s budget allowed. Under sustained AI training workloads, heat couldn’t dissipate efficiently. Samsung’s parts on the 1anm DRAM node missed yield targets — and the timing couldn’t have been worse. The highest-volume AI GPU production cycle in history was just ramping up.
Nvidia qualifies HBM suppliers for each GPU platform independently. Losing HBM3E qualification for H100 and early Blackwell builds (B200/B300) meant Samsung couldn’t participate in the highest-margin memory window of the AI infrastructure boom. It got pushed into lower-margin segments, pricing HBM3E 12-layer products approximately 30% below SK Hynix just to secure whatever allocation it could.
By Q2 2025, Samsung held only 17% HBM share. Counterpoint’s Q3 2025 data showed Samsung at 22%, SK Hynix at 57%. Samsung eventually received HBM3E 12-Hi qualification from Nvidia in November 2025 — but by then SK Hynix had locked in multi-year hyperscaler supply contracts that qualification alone couldn’t unwind.
Here’s the distinction that gets conflated everywhere: Samsung reclaimed total DRAM revenue leadership in Q4 2025. Real, worth noting. But DRAM revenue leadership and HBM-specific revenue leadership are completely separate measurements. In HBM, Samsung still trails at 22% versus SK Hynix’s 57–70%. Conflating the two gives you a misleading picture of the competitive position that actually matters for AI infrastructure.
SK Hynix by the numbers: record profits, sold-out capacity, and a $13 billion expansion bet
SK Hynix reported Q1 2026 operating profit of 37.61 trillion won at a 72% operating margin — a record for any memory company in any quarter. For full year 2025, SK Hynix’s 47.2 trillion won surpassed Samsung’s 43.6 trillion won. Samsung’s memory segment generated approximately 24.9 trillion won in 2025 — meaning SK Hynix earned nearly double Samsung’s memory-only profit.
HBM share sits at approximately 57% per Counterpoint’s Q3 2025 data, climbing toward 70% by TrendForce’s Q1 2026 estimates. All 2026 HBM production is sold out.
SK Hynix has committed ~19 trillion won (~$13 billion) to the new Cheongju M15X mega-fab. SK Group Chairman Chey Tae-won stated publicly in March 2026 that the wafer shortage will likely persist until 2030. M15X is forward-looking — targeting HBM4 and HBM4E ramps. SK Hynix is planning around a multi-year constrained market, not a short cycle.
One thing worth flagging: approximately 42% of SK Hynix’s total revenue comes from a single customer — Nvidia. The dominant position is real, but the financials are heavily exposed to any change in Nvidia’s sourcing strategy. That dependency belongs in any honest assessment of SK Hynix as a long-term supplier.
Is Samsung’s HBM4 recovery bid credible — or is SK Hynix still pulling ahead?
Samsung shipped its first commercial HBM4 on February 12, 2026. SK Hynix had its HBM4 mass-production system in place in September 2025, delivering large volumes of paid samples to Nvidia well before that. The production readiness gap is roughly five to six months, following about one year of sampling lead time.
That lag is evidence the quality gap from HBM3E is not fully closed. Samsung is recovering while SK Hynix continues to scale — those aren’t the same thing.
Nvidia’s Vera Rubin HBM4 allocation reflects the current position: approximately two-thirds to SK Hynix, one-third to Samsung. An improvement from near-zero during peak HBM3E allocation — but not parity.
Samsung’s most concrete win this cycle is the AMD MI400 AI accelerator, committed primarily to Samsung HBM4. That’s the first major AI GPU socket Samsung has fully won this cycle, giving it a foothold independent of Nvidia. Samsung’s internal projection is that HBM sales will more than triple in 2026 — genuine momentum, though tripling still leaves it well behind.
Samsung is also working to reduce its dependency on TSMC‘s CoWoS-L packaging capacity, sold out through 2027. Samsung’s exploration of alternatives to CoWoS packaging maps this in detail.
The more interesting story is HBM4E, expected in 2027. Samsung is the only memory company with an in-house logic foundry. HBM4 and HBM4E require a logic base die fabricated on an advanced node. Samsung can co-design the DRAM dies and the logic base die within a single organisation — no external foundry dependency, no competing for TSMC allocation. No other memory company can do that. The battleground for parity is HBM4E, not HBM4.
How does a near-duopoly market structure keep DRAM prices elevated beyond the HBM cycle?
As Micron’s business chief Sumit Sadana put it: “As we increase HBM supply, it leaves less memory left over for the non-HBM portion of the market, because of this three-to-one basis.”
That’s the crux of it. Each wafer allocated to HBM displaces approximately three wafers of conventional DDR5 output. With HBM3E ASPs running six to eight times conventional DDR5 per gigabyte — and HBM4 at roughly ten times — both manufacturers have every incentive to prioritise HBM. DDR5 supply tightens regardless of DDR5 demand.
TrendForce data shows DRAM contract prices rose 90–95% quarter-over-quarter in Q1 2026, with a further 58–63% increase projected in Q2 2026. NAND prices rose 70–75% in Q1 2026. Dell raised hardware prices 17% on March 30, 2026; HP reported memory costs doubled in a single quarter, now representing 35% of PC build materials.
In a normal DRAM cycle, Samsung could flood supply to gain share. In HBM, that’s not possible — qualification takes 6–18 months per product cycle. That lock-in is what makes this shortage structurally different: the duopoly has pricing power in both HBM and conventional DRAM simultaneously, because wafer capacity is coupled. The DRAM budget impact analysis covers the enterprise cost implications in detail.
The one wildcard in this structure is Micron as the US entrant that could disrupt this duopoly. Micron holds 21–24% HBM revenue share today and has its HBM supply sold out through 2026 and 2027 — but as the sole US-headquartered HBM manufacturer, its trajectory matters for both enterprise procurement options and geopolitical supply chain resilience.
Why are enterprise buyers at the back of the queue — and what does that mean for procurement?
The allocation structure has three tiers, and enterprise buyers are at the bottom.
At the top, hyperscalers — Microsoft, Google, Amazon, Meta — have signed multi-year Long-Term Agreements (LTAs) that lock in supply and pricing. SK Hynix is reportedly receiving direct financing from Big Tech in exchange for committed supply. Google has stationed procurement executives in South Korea. These are not negotiations enterprise buyers can participate in.
Below that, buyers without multi-year LTAs compete for residual allocation at spot pricing. DRAM prices have surged 80–90% this quarter. Lead times have stretched beyond 40 weeks, beyond 58 weeks for some configurations.
For HBM: enterprise buyers can’t purchase from SK Hynix or Samsung at all. HBM is sold exclusively to GPU OEMs — Nvidia, AMD — who integrate it into products. Your constraint is GPU availability and server OEM lead times, not direct memory procurement.
For conventional DRAM and NAND, two things to accept:
Multi-year contracts are now a prerequisite, not a negotiating advantage. Without an LTA, you’re on spot pricing with no predictability.
Infrastructure planning has to run over 18–24 month horizons. Spot-market buying for server memory is a losing strategy. Buyers who delay face higher prices and longer waits simultaneously.
The SK Hynix Q1 2026 earnings call framed it well: “customers are prioritising procurement over price.” Paying more doesn’t move you up the queue when production is fully committed. This is a calendar and relationship problem. For a complete view of the structural forces shaping every layer of this market — from hyperscaler contracts through to enterprise and consumer repricing — see the AI memory crunch overview.
Frequently Asked Questions
Why does SK Hynix have roughly 70% HBM market share when Samsung is a larger company overall?
SK Hynix’s MR-MUF process produces higher TSV yield than Samsung’s TC-NCF method. At 12 stacked dies, the compounding is significant: 75–80% stack yield (SK Hynix 1bnm) versus 60–65% (Samsung 1cnm). SK Hynix was the sole HBM3 supplier to Nvidia’s H100 and used that position to lock in multi-year contracts. Samsung’s HBM3E qualification delays in 2024–2025 then excluded it from the highest-volume Blackwell GPU run. Samsung reclaimed total DRAM revenue leadership in Q4 2025 — real, but a different measurement from HBM-specific leadership.
Did Samsung lose Nvidia as a customer over HBM3E quality issues?
Not entirely. Samsung received HBM3E 12-Hi qualification from Nvidia in November 2025, but missed the bulk of Blackwell volume production. Samsung is now an active HBM4 supplier for Vera Rubin, receiving approximately one-third of allocation versus SK Hynix’s two-thirds. The relationship is recovering, not severed.
Can enterprise buyers negotiate directly with SK Hynix for HBM supply?
No. HBM is sold exclusively to GPU OEMs — Nvidia, AMD. Enterprise buyers access it indirectly through AI accelerator hardware. For conventional DRAM and server memory, buyers can negotiate with distributors and OEMs, but without multi-year LTAs they face spot pricing and lead times beyond 40 weeks.
Will Samsung recover its HBM market share in 2027?
Samsung’s HBM4 programme is live (first shipment February 2026) and HBM sales are projected to more than triple in 2026. TrendForce estimates Samsung’s share could surpass 30% in 2026. Parity with SK Hynix isn’t expected before HBM4E. Samsung’s best path is foundry integration: co-designing DRAM and logic base die in-house is a differentiator no other memory company can replicate.
What is the difference between HBM3E and HBM4, and who leads in each generation?
HBM3E is the current volume product: 12-Hi stacking, ~1.2 TB/s per stack; SK Hynix dominates at 57–70%. HBM4 doubles the interface bus to 2,048 bits and requires a logic base die on an advanced process node. SK Hynix sampled HBM4 roughly one year before Samsung’s first commercial shipment and holds ~70% of Nvidia Vera Rubin allocation. The next battleground for parity is HBM4E in 2027.
What does MR-MUF vs TC-NCF mean in practical terms for HBM quality?
MR-MUF (SK Hynix): all 12 dies bond in a single thermal cycle — fewer voids, higher yield; 75–80% stack yield on the 1bnm node. TC-NCF (Samsung): non-conductive film applied layer by layer, historically slower and lower yield; “Advanced TC-NCF” narrowed the gap; 60–65% yield on the 1cnm node. The result: SK Hynix produces more usable stacks per wafer with fewer failures under sustained thermal load — exactly the condition during AI training.
Why is the HBM memory shortage described as “structural” rather than cyclical?
Manufacturers can’t add HBM supply without Nvidia or AMD qualification — 6–18 months per cycle. The three-to-one wafer trade ratio means adding HBM supply simultaneously tightens conventional DRAM supply. IDC characterises this as a permanent strategic reallocation of global wafer capacity, with supply growth at 16% year-on-year versus 20–30% historical norms.
What is SK Hynix’s Nvidia revenue concentration risk?
Approximately 42% of SK Hynix’s total revenue comes from Nvidia alone. The dominant position is real, but the financials are disproportionately exposed to any shift in Nvidia’s sourcing strategy, GPU architecture changes, or demand softening. Worth factoring into any assessment of SK Hynix as a durable long-term supplier.
What is the Cheongju M15X fab and why does it matter?
The Cheongju M15X is SK Hynix’s new mega-fab in South Korea — a ~$13 billion investment targeted at HBM4 and HBM4E production ramps. SK Group Chairman Chey Tae-won’s March 2026 statement that the wafer shortage will likely persist until 2030 shows how SK Hynix is planning: around a multi-year constrained market, not a short-term cycle.
