The HBM4 handover: who actually supplies the AI memory in 2026
SK hynix still leads high-bandwidth memory, Micron has overtaken Samsung on some measures, and the HBM4 qualification race for NVIDIA's Rubin platform resets the board. What the share numbers mean and what they don't.
Roughly two Korean companies decide how fast the world’s AI infrastructure can be built. That is not hyperbole — it is a description of the high-bandwidth memory market, and it is the reason the Korean stock index behaved like a leveraged AI bet in 2026.
The competitive picture inside that market changed meaningfully this year. Here is what actually moved.
First, the number that isn’t one number
If you search for HBM market share you will find figures that contradict each other by twenty percentage points. All of them are being quoted in good faith. Consider a representative sample from 2026 coverage:
| Source framing | SK hynix | Samsung | Micron |
|---|---|---|---|
| Q2 2025 share | 62% | 17% | 21% |
| Q1 2026 (Counterpoint) | 58% | — | — |
| More recent range estimates | 50–55% | 35–40% | 5–10% |
These cannot all describe the same quantity, and they don’t.
What is consistent across every source is the direction of travel: SK hynix still leads, and its lead is narrowing from an extraordinarily high base. Roughly 69% a year ago. Around 58% in Q1 2026 on Counterpoint’s numbering. That is still dominance. It is also a company giving up share every quarter, which is a different story from the one told in 2024.
The genuinely new fact: Micron passed Samsung
The headline most people missed is not about SK hynix at all. On several 2026 estimates, Micron overtook Samsung in HBM.
That would have been an implausible claim two years ago. Samsung is the largest memory manufacturer on earth by capacity. But HBM is not commodity DRAM — it is a stacking, packaging, thermal and yield problem as much as a lithography one, and Samsung’s difficulties qualifying HBM3E with the dominant accelerator customer cost it a generation of position while Micron executed cleanly.
The lesson generalises: in HBM, capacity is necessary and nowhere near sufficient. Qualification is the gate.
What HBM4 changes
The market is growing from roughly $38bn in 2025 to $58bn in 2026. HBM3E still accounts for about two-thirds of 2026 shipments; HBM4 ramps through the second half.
HBM4 is not a routine increment. Two changes matter:
The interface widens dramatically. More parallel lanes at lower per-lane rates — better bandwidth per watt, harder signal integrity and packaging.
The base die becomes a logic problem. The layer at the bottom of an HBM4 stack moves closer to a custom logic die, which drags foundry-grade capability into what used to be a pure memory competition. That is why HBM4 positions are not simply inherited from HBM3E: the required skill mix has changed, and companies with strong logic and packaging arms have a route back in.
The capacity race underneath
Both Korean suppliers are spending against the same bet, that AI memory demand outruns supply through 2027:
- Samsung is targeting roughly 50% capacity expansion in 2026.
- SK hynix has raised planned infrastructure investment to more than four times its previously announced figure.
Numbers of that shape carry a specific risk. HBM capacity is not fungible — it is built for a generation and a customer, with committed advanced-packaging lines behind it. If accelerator demand slows while three suppliers are mid-expansion, the industry gets a classic memory glut, except in a product with much higher fixed costs. The memory industry has run this experiment before, repeatedly, and it has never ended gently.
Why this is a Korea story and not just a semiconductor story
Because of concentration. Samsung Electronics and SK hynix together make up a very large share of the KOSPI, and HBM economics drove their 2026 earnings expectations. When the index gained about 92% in the first half of the year and then fell some 40% from its June peak, that was — in substance — the market repricing this one product cycle twice in six months. We covered the mechanics in the 2026 crash explainer.
For anyone holding a Korea ETF, the practical implication is blunt: you own the HBM cycle, whether or not you meant to. Our guide to buying Korean stocks from abroad sets out how to see that concentration in your own holdings before you size a position.
Figures current as of 12 August 2026, sourced to the outlets and company disclosures listed below. Where trackers disagree, we have shown the range rather than selecting one.
Useful links & tools
Official portals and primary data sources for this topic. Opens in a new tab.
- SK hynix newsroom — market outlook and HBM announcements — First-party product and capacity announcements, in English.
- Samsung Semiconductor newsroom — Product launches, process-node and packaging announcements.
- DART — Korean company filings — Quarterly results and capex disclosures for both Korean suppliers, in primary form.
- Counterpoint Research — DRAM and HBM market share tracker — One of the more transparent public share trackers; note its methodology before quoting it.
- Korea Customs Service — trade statistics — Semiconductor export values by month — the fastest public read on the cycle.
Frequently asked questions
Why do published HBM market-share figures disagree so much?
Because they measure different things. Some trackers count revenue, some count bit shipments, some count wafer starts. Some report a single quarter, some a trailing year. HBM3E and HBM4 carry very different prices per bit, so a revenue-share number and a volume-share number can diverge by twenty points and both be correct. When you see an HBM share figure, always establish three things: revenue or volume, which quarter, and which generations are included.
What actually is HBM, in one paragraph?
High-Bandwidth Memory is DRAM stacked vertically into a cube, connected by through-silicon vias, and placed on the same package as the processor. Instead of data crossing a long, narrow path to distant memory modules, it crosses a very short, very wide one. For AI training and inference — which are memory-bandwidth-bound far more than they are compute-bound — this is the difference between an accelerator running at capacity and one waiting for data.
Is HBM4 a bigger jump than previous generations?
Structurally, yes. HBM4 widens the interface substantially and, critically, changes the base die — the logic layer at the bottom of the stack — into something closer to a custom logic component, which pulls foundry capability into what was previously a pure memory problem. That is why the competitive order in HBM4 is not guaranteed to match HBM3E: it rewards a different mix of skills.
Does this mean Samsung is losing?
It means Samsung lost the HBM3E generation’s early lead and is trying to recover position in HBM4. It is the world’s largest memory manufacturer by overall capacity and it is expanding aggressively. A company can simultaneously be third in one high-value product line and structurally central to the industry. Both things are true here.
How does this connect to the KOSPI?
Directly and almost mechanically. Samsung Electronics and SK hynix dominate the Korean index, and HBM economics drove their 2026 earnings expectations. The index’s near-doubling and subsequent crash were, in substance, a repricing of this one product cycle — see our 2026 crash explainer.
What would change the picture fastest?
A qualification decision. HBM supply is sole-sourced or dual-sourced per accelerator platform, and being designed into a flagship product locks in volume for years. Watch for qualification announcements around NVIDIA’s Rubin generation and for custom-accelerator programmes at the large cloud providers, which increasingly specify their own memory partners.
Sources & further reading
Every figure in the key takeaways is numbered to the source it was read from. Sources marked primary are the statistics office, central bank, exchange, regulator or filing itself.
- 12026 Market Outlook: SK hynix's HBM to fuel AI memory boom — SK hynix newsroom
- 2SK hynix holds 62% of HBM, Micron overtakes Samsung, 2026 battle pivots to HBM4 — Astute Group, 2026
- 3Global DRAM and HBM market share, quarterly — Counterpoint Research
- 4Samsung and SK Hynix to scale up memory production capacity in 2026 — Data Center Dynamics, 2026
- 5SK hynix, Samsung: HBM4 readiness, 1c DRAM scaling and the 2026 capacity race — Silicon Analysts, 2026
Read next
Korea's 2026 market crash, explained: what broke, what it cost, and where the KOSPI stands now
The KOSPI roughly doubled, then fell 44% from its all-time high of 9,385.59 on 19 June. Here is what actually caused it — leveraged single-stock ETFs, margin debt and a two-stock index — and where the market stands in August 2026.
How to buy Korean stocks from abroad in 2026: ADRs, ETFs and direct access, compared
Four realistic routes into Korean equities for a non-resident investor — ADRs, US-listed ETFs, local-listed ETFs and a direct KRX account — with the costs, tax treatment, currency exposure and access frictions of each.
Korea's 2026 visa overhaul: Top-Tier, E-7-M and F-2-R explained
Korea opened up its skilled-migration system in 2026: an expanded Top-Tier visa, an E-7-1 experience waiver, a new manufacturing route, a regional residence track, and e-visa processing. What each one requires and who should use which.