Hot Chips 2026: Samsung brings the base die, SK hynix brings the packaging
Samsung Electronics and SK hynix will present next-generation HBM work at Hot Chips 2026 at Stanford on August 23–25, Samsung on the base die and SK hynix on advanced packaging, as forecasts point to Samsung taking a 41% bit-shipment share next year against SK hynix's 39%
The three lines
- Samsung presents HBM base die; SK hynix presents HBM advanced packaging
- Samsung emphasizes vertically integrated custom memory; SK hynix emphasizes open standards and ecosystem
- HBM bit-shipment share forecast: SK hynix 48% and first this year, Samsung 41% vs SK 39% next year
Key questions
- What is Hot Chips
- An annual chip-architecture conference held at Stanford. It is not a product launch venue — companies disclose internal design details and take questions from engineers, which is why the industry reads it as a roadmap signal. The 2026 edition runs August 23–25 local time.
- Base die versus packaging — what is the difference
- HBM stacks memory dies vertically to widen bandwidth. The base die sits at the bottom of that stack, consolidating signals to the GPU and taking on some logic; packaging is the back-end process of bonding the layers and getting heat out. Samsung is presenting on making the bottom smarter; SK hynix on improving how the stack is assembled. Both are real bottlenecks.
- Will the share lead actually flip
- It is a forecast, not a result. On HBM bit shipments SK hynix holds first place at 48% this year, with Samsung projected at 41% against SK hynix's 39% next year. Two caveats: this measures bits, not revenue, and HBM customers are concentrated enough that a single qualification decision can move several points at once.
Yesterday this page described the emerging post-HBM split: Samsung stacking memory onto the GPU, SK hynix extending through NAND. That framing now gets a venue. Hot Chips 2026 runs August 23–25 local time at Stanford University, and both Korean memory makers are on the program.
They are speaking at the same conference on different subjects. The difference is the story.
1. The bottom of the stack, or the way it is bonded
| Company | Hot Chips 2026 topic | Where it sits |
|---|---|---|
| Samsung Electronics | HBM base die | The bottom die — signal consolidation and some logic |
| SK hynix | HBM advanced packaging | Back-end — bonding layers and thermal extraction |
HBM widens bandwidth by stacking memory dies vertically. That structure has two chokepoints. One is the bottom: signals from every layer converge at the base die before reaching the GPU, so how much function can be packed in there sets much of the next generation's performance. The other is assembly: as layers multiply, heat gets trapped and bond interfaces accumulate, so packaging effectively determines the stacking ceiling.
Samsung is bringing the first; SK hynix the second. The conference will not adjudicate which choice is right. But the fact that each company is showing the part it is confident in locates the center of gravity of its roadmap.
2. Integration versus standards
At FMS earlier this month the two firms laid out next-generation AI memory roadmaps side by side, and the contrast there is likely to repeat.
Samsung emphasizes custom memory built on vertical integration — using design, fabrication and packaging under one roof to tailor HBM to a specific customer's requirements. The base-die talk fits that exactly, since the base die is where customization has the most room.
SK hynix emphasizes open standards and ecosystem breadth — keeping specifications wide enough that many customers and partners build on the same footing. A packaging talk aligns with process improvements that benefit the whole line rather than one account.
Each strategy wins under different market conditions. Customization pays when a handful of enormous buyers dominate; standardization pays as the buyer base broadens. Which way the AI accelerator market develops is, in effect, the referee.
3. The forecast, and its conditions
| HBM bit shipments | This year | Next year (forecast) |
|---|---|---|
| SK hynix | 48% (first) | 39% |
| Samsung Electronics | — | 41% (first) |
Two cautions attach. This measures bits, not revenue — generations carry different prices per bit, so a revenue ranking can differ. And these are research-firm estimates that vary between firms.
More importantly, HBM share does not behave like a consumer market. Buyers are concentrated in a small number of AI accelerator vendors, so one qualification decision can move several points at once. A projected 41%-to-39% gap in a market like that is less a result than an unsigned contract.
4. What remains unverified
Session specifications are not published before the conference. What is confirmed today is the topics; the actual numbers — bandwidth, stack height, power — arrive in late August. Share forecasts are estimates, and the "vertical integration versus open standard" contrast is this publication's reading of the topics and public roadmaps, not language either company used.
Background on the technology is in "HBM explained: why stacked memory is AI's bottleneck", on the next-generation split in "What comes after HBM", and on the lag between capex and prices in "How long semiconductor capex takes to reach prices".
Sources
- Daum/Electronic Times — Samsung and SK hynix to present next-gen HBM at Hot Chips
- Financial News — Samsung vs SK hynix on post-HBM leadership at FMS
- The Fact — Samsung stacks, SK connects: who leads post-HBM
- SK hynix Newsroom — 2026 market outlook: the HBM-led memory supercycle
- Nate/The Fact — Biz Talk: the post-HBM leadership race