Samsung Electronics used the Future of Memory and Storage (FMS) 2026 expo in Santa Clara on August 4 to unveil zHBM, a memory architecture that stacks High Bandwidth Memory directly on top of an AI accelerator instead of placing it beside the chip on an interposer. According to Samsung’s official announcement, a next-generation interface system built around zHBM is projected to deliver roughly eight times the performance per GPU of HBM5, the generation that hasn’t even reached commercial shipment yet.
The reveal landed at a pointed moment. DRAM and HBM prices have been climbing all year on AI server demand, Samsung has slipped to third place in HBM market share behind SK hynix and Micron, and Nvidia’s Rubin GPU platform is already locking in HBM4 supply deals that mostly favor Samsung’s rivals. zHBM is Samsung’s attempt to change the conversation from “who ships the most HBM4 today” to “who owns the architecture that comes after HBM5.” Whether that bet pays off depends on manufacturing steps nobody has proven at volume yet.
What Samsung actually announced at FMS 2026
Kim Kyoung-ryun, vice president of Samsung’s DRAM Development Office, presented the zHBM concept at FMS 2026, according to SBS English. The pitch is straightforward on paper: conventional HBM sits next to a GPU die on a silicon interposer, and data travels sideways across that interposer before it reaches the compute logic. zHBM flips the geometry, stacking memory dies vertically in the Z-axis directly above the accelerator, cutting the physical distance signals have to travel.
Samsung’s DS Division CTO and president, Jae-hyuk Song, previewed the underlying idea back in February at Semicon Korea, months before the FMS reveal gave it a name and a performance target. “We plan to develop zHBM technology, which places HBM itself on the Z-axis,” Song said, according to Sedaily. At that point Samsung was framing the comparison against HBM4, not the newer HBM5 baseline used in August. “It has shown the potential to quadruple both bandwidth and power efficiency compared to HBM4,” Song added, per the same Sedaily report.
By the time FMS 2026 rolled around, Samsung had a bigger number and a tougher comparison target. The company said thermal resistance under zHBM drops to somewhere between 10% and 25% of HBM5 levels, and performance-per-watt improves roughly threefold over a conventional HBM-plus-interposer layout, per SBS English’s report on the Santa Clara presentation. Those are Samsung’s own projections, not third-party lab measurements, and the company has not disclosed a firm date for when zHBM chips would actually ship.
Why stacking memory on top of the GPU matters
The industry has been circling what engineers call the memory wall for years: compute performance on AI accelerators has scaled faster than the bandwidth available to feed those chips with data. HBM was itself a response to that problem, but even HBM4 still relies on a 2.5D layout, memory dies sit flat, side by side with the GPU, connected through an interposer. Every trip across that interposer costs energy and adds a small amount of latency, and as GPUs pack in more compute, wider and wider interposer connections are needed just to keep pace.
A Samsung executive identified as Park described the company’s direction in blunter terms to trade press, framing the goal as folding the accelerator and the memory stack into a single physical structure rather than two components wired together. The industry shorthand for this is true 3D integration, as opposed to the 2.5D approach every current HBM generation uses. Getting there requires hybrid bonding, a technique that fuses copper pads from two separate wafers or dies directly together without solder bumps, at pitch tolerances tighter than 10 microns.
That manufacturing step is the real story behind the story. Analyst firm Futurum Group has flagged hybrid bonding at sub-10-micron pitch as the hardest part of the zHBM roadmap, arguing that Samsung’s announcement reads more like a statement of direction than a near-term product, according to Futurum Group’s analysis. Every major memory maker, not just Samsung, is chasing the same hybrid-bonding milestone for different products, which is why zHBM should be read as a preview of where the whole HBM roadmap is heading rather than a Samsung-only breakthrough.
zHBM by the numbers: HBM4, HBM4E, HBM5 compared
Samsung’s roadmap disclosures at FMS 2026 double as a status update on where each HBM generation actually stands. The company said it became the first in the industry to mass-produce HBM4, using 1c DRAM and 4nm base-die technology, starting February 6, 2026. It followed that in May 2026 by shipping HBM4E samples to global customers, ahead of a generation most third-party roadmaps still expect to ramp through 2027. HBM5 itself is still in development and, per Samsung’s disclosures, will be the first HBM generation to use Gate-All-Around transistors on a 2-nanometer foundry process.
| Generation | Status as of August 2026 | Key claim or spec |
|---|---|---|
| HBM3E | Shipping in volume, primary generation in current AI GPUs | Established baseline for current-gen accelerators |
| HBM4 | Mass production started Feb. 6, 2026 (Samsung) | Built on 1c DRAM and 4nm base-die process |
| HBM4E | Samples shipped to customers May 5, 2026 (Samsung) | Enhanced HBM4 follow-on, broader ramp expected 2027 |
| HBM5 | In development, not yet commercially shipping | First HBM generation on 2nm Gate-All-Around process |
| zHBM | Concept unveiled Aug. 4, 2026 at FMS, no ship date given | Claimed ~8x performance per GPU vs. HBM5, ~3x perf-per-watt |
Read the table in order and the pattern is clear: Samsung is racing to close gaps on multiple fronts simultaneously, current-generation HBM4 volume, next-generation HBM4E samples, and a longer-horizon architecture bet in zHBM, all inside the same twelve months. That’s an aggressive cadence even by memory-industry standards, and it mirrors what’s happening across the wider chip stack we’ve covered before in our look at how Nvidia has compressed its own AI model release cycle to keep pace with competitors.
zNAND-O and the rest of Samsung’s FMS 2026 lineup
zHBM wasn’t the only 3D-stacking concept Samsung showed off in Santa Clara. Lee Jin-yeop, vice president of Samsung’s Flash Development Office, presented zNAND-O, a stacked-NAND approach aimed at on-device AI workloads rather than data-center GPUs. Samsung said internal testing on a 120-billion-parameter GPT-class model showed zNAND-O matching the performance of a conventional DRAM-based server setup at roughly one-sixth the cost, per SBS English’s coverage of the presentation.
Samsung also used the show to detail a tenth-generation BV-NAND design with more than 400 stacked layers and a new high-end enterprise SSD line, positioning the announcements together as a single “next-gen 3D memory vision” rather than a series of unrelated product reveals. The company’s framing, unusual for a memory maker that historically talked in bandwidth and capacity numbers, leans hard into vertical integration as the organizing idea behind its entire 2026-2027 product roadmap.
Samsung’s market position: why the urgency now
Samsung isn’t unveiling ambitious memory architecture out of pure engineering enthusiasm. It’s trying to reset a competitive position that’s slipped over the past two years. SK hynix has held the largest share of the HBM market through most of 2026, reported at roughly 58% in the first quarter by Counterpoint Research data cited by Tom’s Hardware, down from a prior high near 69% as competitors caught up. More striking for Samsung specifically: Micron overtook it for the number-two HBM position in several 2026 share tallies, pushing Samsung into third place in a market it once helped pioneer.
The stakes tied to Nvidia’s next GPU generation make that ranking even more consequential. UBS analysts have projected SK hynix will hold roughly 70% of the HBM4 supply feeding Nvidia’s Rubin platform in 2026, according to reporting on SK hynix’s own market outlook. That leaves a shrinking slice for Samsung and Micron to split on the highest-volume AI GPU program of the year, which helps explain why Samsung is pushing HBM4, HBM4E, and zHBM concepts out the door in close succession rather than waiting for a cleaner, single-generation launch.
| Company | Approx. HBM market share (2026) | 2026 roadmap status |
|---|---|---|
| SK hynix | ~58-62% (Counterpoint Research, various 2026 quarters) | Leading HBM4 supplier for Nvidia Rubin, UBS projects ~70% Rubin HBM4 share |
| Micron | ~21% (2026 estimates) | Overtook Samsung for #2 share, HBM4 in mass production, HBM4E targeted 2027-2028 |
| Samsung | ~17% (2026 estimates) | HBM4 mass production Feb. 2026, HBM4E samples May 2026, zHBM unveiled Aug. 2026 |
None of these share figures are precise to the decimal across sources, different research houses report slightly different quarters and methodologies, but the direction is consistent everywhere it’s measured: SK hynix leads by a wide margin, Micron has closed in on second place, and Samsung is defending third. That context matters for reading zHBM correctly. It’s less a triumphant new product launch and more a signal that Samsung wants to compete on the next battlefield rather than keep fighting for share on the current one.
The DRAM shortage that’s funding this arms race
All three companies are investing heavily in HBM at the same time DRAM pricing overall has gone through its steepest run-up in years. JPMorgan Global Research said in August 2026 that DRAM prices were on pace to rise more than 400% between the start of 2024 and the end of 2026, a figure driven almost entirely by AI server demand pulling capacity away from consumer and enterprise buyers. TrendForce’s more granular quarterly tracking backs that up: the research firm forecast conventional DRAM contract prices rising 13% to 18% quarter-over-quarter in the third quarter of 2026, after an even steeper 58% to 63% quarterly jump it recorded for the second quarter, per TrendForce’s pricing update.
Those price spikes haven’t stayed confined to server hardware. We’ve already tracked how consumer RAM prices climbed as AI demand siphons memory capacity, how Apple’s M6 and M5 Ultra chips pushed configurations up to 512GB of unified memory even as component costs rose, and how a laptop maker recently reintroduced DDR4 configurations on new RTX 5070 laptops specifically to dodge DDR5 pricing. For a broader look at where the chip market is headed next, see our hardware and chips coverage. Every fab cycle Samsung, SK hynix, and Micron devote to HBM output is capacity that isn’t going toward conventional DDR5 production, which is part of why the shortage keeps showing up on both the data-center and consumer sides of the market at the same time.
How this connects to Nvidia’s Rubin and the broader AI buildout
Nvidia’s Rubin platform is the backdrop against which all of this HBM competition is playing out. Whichever memory supplier locks in the largest share of Rubin’s HBM4 and HBM4E allocation effectively locks in a multi-billion-dollar revenue stream tied to Nvidia’s release cadence, which is a big part of why UBS’s 70% share projection for SK hynix reads as more than a routine analyst estimate. That HBM4 supply relationship is the same one we detailed in our coverage of HBM4 yield rates powering Nvidia’s Rubin platform, and zHBM’s entire premise is built on top of that existing HBM4-to-Rubin pipeline. It’s the number Samsung and Micron are both trying to chip away at with faster qualification cycles and, in Samsung’s case, an entirely different architecture pitch in zHBM.
The memory side of the AI buildout gets less attention than GPU launches themselves, but it’s arguably the tighter bottleneck. GPU compute has scaled through new process nodes and architecture generations relatively predictably. HBM supply has not kept pace, which is exactly the gap zHBM is designed to address structurally rather than just incrementally. Samsung’s framing, and Chosun Biz’s reporting on the company’s roadmap backs this up, treats zHBM less as a standalone product and more as the architecture Samsung expects HBM5E and later generations to eventually adopt.
What analysts and competitors are saying
Samsung has been consistent in its own public statements about zHBM’s potential, even as the performance comparison shifted between February and August. In its official FMS 2026 materials, Samsung stated plainly: “A next-generation interface system incorporating zHBM is expected to deliver approximately eight times the performance of HBM5,” according to the company’s newsroom announcement. The same announcement traced the company’s momentum through the year, noting: “Following the industry’s first mass production of HBM4 using its 1c DRAM and 4-nanometer (nm) base die technologies in February, Samsung was the first to begin shipping HBM4E samples to global customers in May.”
Outside analysts have been more cautious than Samsung’s own messaging. Futurum Group’s assessment treated the hybrid-bonding manufacturing step as the real gating factor for zHBM reaching production, rather than the architecture concept itself, a distinction that matters because Samsung has not committed to a mass-production timeline. Chosun Biz’s reporting placed HBM5 and HBM5E commercialization around 2028 in most third-party roadmaps, which implies zHBM-style stacking is more likely to arrive as part of that generation transition than as an earlier standalone product.
Competitive comparison: Samsung’s bet versus the field
SK hynix’s public strategy through 2026 has centered on scaling what it already does well: hybrid bonding for its own next-generation HBM packaging, and volume. The company has emphasized manufacturing execution and customer lock-in with Nvidia over architectural reinvention, a strategy that’s working given its share lead. Micron’s approach has focused on closing the qualification gap, moving HBM4 into mass production and pointing toward HBM4E in the 2027-2028 window without publicly previewing a post-HBM5 architecture the way Samsung just did.
Samsung’s zHBM reveal is the outlier in that field: a bet that architecture, not incremental generation-over-generation scaling, is where the next real advantage sits. That’s a higher-risk, higher-reward strategy. If hybrid bonding at sub-10-micron pitch proves manufacturable at volume on Samsung’s timeline, zHBM could leapfrog the field the way Samsung’s press materials suggest. If it slips, and analyst commentary suggests that’s the more likely near-term outcome, Samsung spends 2026 and 2027 still fighting SK hynix and Micron on HBM4E execution while its architectural bet matures on a slower track.
Historical context: how we got from HBM1 to zHBM
HBM has moved through roughly one major generation every two to three years since its commercial debut in the mid-2010s. HBM2 brought higher capacity and bandwidth into mainstream GPU products later in that decade, and HBM2E extended that generation with speed and capacity bumps. HBM3 arrived alongside the first big wave of generative-AI accelerator demand, and HBM3E became the workhorse generation powering most of the AI GPUs shipping today. HBM4 is the generation now moving through qualification and early mass production across all three major suppliers, with HBM4E following close behind rather than waiting for a full multi-year gap the way earlier transitions did.
What’s different about the HBM4-to-HBM5 stretch is the compressed timeline. Where earlier HBM generations shipped years apart with clear daylight between them, Samsung alone moved from HBM4 mass production to HBM4E samples to a zHBM architecture preview inside a single calendar year. That compression reflects how much pressure AI accelerator demand has put on the entire memory roadmap, there’s no longer room for the industry to treat memory generations as a slow, predictable cadence.
What this means for AI hardware buyers and data center planners
For hyperscalers and enterprise buyers planning GPU cluster procurement into 2027 and 2028, zHBM is a signal worth tracking rather than a spec to plan around today. Nothing about Samsung’s announcement changes what’s available to order right now, HBM4 and HBM4E remain the near-term products actually shipping, and pricing on those generations is still climbing under the broader DRAM shortage described above. Buyers locking in multi-year supply agreements should expect that pressure to persist rather than ease, regardless of how zHBM’s development plays out.
What zHBM does change is the long-range planning conversation. If Samsung’s architecture bet pans out even partially, it reshapes which supplier data center operators want to anchor multi-year contracts with beyond the current HBM4E cycle. That’s a decision most large buyers won’t need to make for another year or two, but the FMS 2026 reveal is the kind of announcement that starts showing up in supplier-diversification conversations well before any zHBM chip actually ships.
Predictions: what happens next
Based on the roadmap disclosures, analyst commentary, and market-share trends gathered here, a few outcomes look likely over the next 12 to 24 months. First, expect Samsung to keep repositioning its zHBM performance claims as HBM5 development firms up, the shift from a 4x comparison against HBM4 in February to an 8x comparison against HBM5 in August suggests the company will keep recalibrating the headline number as the underlying baseline changes. Second, DRAM and HBM pricing is unlikely to ease meaningfully before 2027 given TrendForce’s and JPMorgan’s current trajectories, so buyers should plan for continued cost pressure rather than a near-term correction.
Third, SK hynix’s HBM4 lead tied to Nvidia’s Rubin platform is likely to persist through at least 2026, given UBS’s share projection and SK hynix’s own manufacturing head start, meaning Samsung’s near-term competitive relief will have to come from HBM4E execution rather than zHBM. Fourth, hybrid-bonding manufacturing progress, not marketing announcements, will be the real signal to watch for whether zHBM becomes a real product or stays a roadmap concept, and that progress is more likely to surface first in HBM5E disclosures than in a standalone zHBM launch. Fifth, expect Micron and SK hynix to respond with their own vertical-stacking architecture previews within the next year, since none of the three suppliers can afford to let a rival own the “what comes after HBM5” narrative uncontested.
Frequently asked questions
What is Samsung’s zHBM?
zHBM is a memory architecture Samsung unveiled at FMS 2026 that stacks High Bandwidth Memory vertically on top of an AI accelerator, rather than placing it beside the chip on a silicon interposer the way current HBM generations do.
When was zHBM announced?
Samsung presented zHBM on August 4, 2026, at the Future of Memory and Storage expo in Santa Clara, California, though the company had previewed the underlying stacking concept in February 2026 at Semicon Korea.
How much faster does Samsung claim zHBM is?
Samsung’s FMS 2026 materials say a next-generation interface system built around zHBM could deliver roughly eight times the per-GPU performance of HBM5, with about three times the performance-per-watt of a conventional HBM-plus-interposer layout.
Is zHBM available to buy?
No. Samsung has not disclosed a mass-production or shipping date for zHBM. HBM4 and HBM4E remain the generations currently shipping to customers as of August 2026.
How does zHBM differ from HBM4 and HBM5?
HBM4 and HBM5 are still 2.5D designs, with memory dies placed beside the GPU on an interposer. zHBM describes a true 3D approach that stacks memory directly above the accelerator using hybrid bonding, a manufacturing step not yet proven at commercial volume.
Why does Samsung’s HBM market share matter here?
Samsung has fallen to roughly third place in HBM market share behind SK hynix and Micron in 2026 estimates. zHBM is widely read as part of Samsung’s push to compete on next-generation architecture rather than continue fighting for share on current HBM4 volume.
Does zHBM affect current DRAM and HBM prices?
Not directly. Current price increases are driven by AI server demand outpacing DRAM and HBM supply across all three major manufacturers, a trend TrendForce and JPMorgan both expect to continue regardless of when zHBM ships.
Will competitors build something similar to zHBM?
Very likely. Hybrid bonding and vertical memory stacking are industry-wide manufacturing goals, not a Samsung exclusive, so SK hynix and Micron are expected to disclose their own stacking roadmaps as HBM5 development progresses.