Infrastructure

Samsung's New Memory Roadmap: Not Buyable Yet

Samsung unveiled three next-gen 3D memory concepts at FMS 2026: zHBM, zNAND-O, and a 400-plus-layer V10 BV-NAND. Here is what ships, what is a vision model, and what it means for AI infrastructure buyers in the EU and UK.

InfrastructureBy Servola Tech Desk2026-08-084 min read

AI-assisted, edited by humans. | Editor-in-Chief: Leon Soliman | Editorial standards

Samsung's New Memory Roadmap: Not Buyable Yet

Key takeaways

Samsung's zHBM concept targets about 8x the performance and more than 10x the memory density of HBM5, but it was shown at FMS 2026 as a vision model with no order date, so it changes nothing about 2026 memory costs yet.
V10 BV-NAND packs more than 400 layers using wafer bonding and raises density by roughly 58 percent over the prior V9 generation, making it Samsung's first bonded-NAND architecture and a fully specified chip rather than a slide.
Samsung already holds about 34 percent of data center memory revenue, ahead of SK Hynix's 29.8 percent and Micron's 17.9 percent, and HBM alone is projected to consume around a quarter of all DRAM wafer output in 2026.
Independent analysis flags thermal extraction, moving roughly 1 kilowatt of heat through a bonded memory stack, as the hardest unresolved production problem, which means qualification and shipping dates, not the spec sheet, are what buyers should track.
European and UK buyers negotiating multi-year AI infrastructure or cloud-capacity contracts in 2026 should price today's memory shortage into those terms now rather than assume a roadmap unveiling will loosen supply within the contract's first year.

Three New Names Built on One Bonding Technique

At the Future of Memory and Storage (FMS) 2026 conference in Santa Clara, Samsung DRAM Design Team vice president Kyungryun Kim and Flash Product and Technology executive vice president Jin-Yub Lee presented three new memory architectures, all built on advanced wafer-bonding technology. zHBM stacks high-bandwidth memory directly on top of an AI accelerator chip instead of placing it beside the processor, and Samsung's own projections put a next-generation interface system using zHBM at roughly 8x the performance of HBM5, more than 10x the memory density, a threefold gain in energy efficiency, and better than a 50 percent cut in thermal resistance.

zNAND-O is a high-performance NAND design built on Samsung's V-NAND line, offered in four- and eight-layer versions and aimed at edge AI workloads that need low latency and real-time processing. V10 BV-NAND is the concrete product of the three: more than 400 layers, built with wafer bonding rather than a conventional single-stack etch, and about 58 percent denser than Samsung's prior V9 generation, with better read, write, and I/O performance.

A Vision Model and a Shipping Chip Are Different Claims

Samsung itself showed zHBM and zNAND-O as concept models, presented alongside products it already ships: HBM4E samples, an HBM5 preview, LPDDR5X-PIM, and the enterprise drives PM1763 and BM1773. Independent industry analysis from Futurum Group sharpens the distinction: V10 BV-NAND is best read as productization of a bonded-NAND concept that rivals including Kioxia, SanDisk, and YMTC have also pursued, with the real competitive edge sitting in bonding-interface yield and alignment precision, not the layer count alone.

Futurum analyst Brendan Burke names three unresolved production hurdles before any of this ships at scale: hybrid-bonding pitch scaling below 10 microns at full DRAM-stack yield, known-good-die testing before an irreversible bond, and thermal extraction, pushing roughly 1 kilowatt of heat through a stacked memory die without degrading data retention. Burke calls the thermal problem the tallest of the three, and it is a materials-science constraint that does not move faster because a roadmap says it should.

The Supply Picture This Does Not Change in 2026

None of this alters the memory market buyers are paying into this year. Samsung already holds about 34.0 percent of data center memory revenue, ahead of SK Hynix at 29.8 percent and Micron at 17.9 percent as of the fourth quarter of 2025, and HBM alone is projected to absorb around a quarter of total DRAM wafer output in 2026.

Analysts tracking the category put the data center off-chip memory market at roughly $17.1 billion in 2025, growing toward $96.8 billion in 2026 and an estimated $260.5 billion by 2030. That is the backdrop an AI buildout has been operating inside through 2025 and 2026: a memory-supply-constrained market tight enough to shape which cloud and chip vendors can promise delivery dates at all, and a concept-model unveiling at a trade show adds no wafer to that supply this year.

What This Means for a European or UK Infrastructure Budget

For an EU or UK enterprise buyer negotiating GPU capacity, a server refresh, or a multi-year cloud contract in 2026, the practical read runs against what a roadmap slide implies. This is not a sign that memory costs are about to ease; it is a supply-side data point showing the vendor with the largest share of data center memory revenue is still years from shipping the technology it says will relieve the bottleneck.

That makes the current shortage the planning baseline for the next one to two budget cycles. The sensible move is to lock in supply and pricing commitments now, while the constraint is known and priced, rather than delaying a contract on the assumption that zHBM-class density gains will show up inside this year's or next year's server refresh.

Reading the Next Year of Memory Announcements

The pattern worth tracking going forward is qualification dates and named customer deployments, not spec-sheet multiples. HBM4E's shipping samples and V10 BV-NAND's stated 400-plus layers are claims with a manufacturing process behind them today; the 8x and 10x figures attached to zHBM describe a target for a system that does not yet exist as a shipping part.

Treat every future memory update the same way: separate what a company is delivering this quarter from what a vision model promises for a generation still working through pitch scaling, die testing, and thermal limits that no press release can shortcut.

Frequently asked questions

What did Samsung actually announce at FMS 2026?

Samsung presented three memory architectures at the Future of Memory and Storage conference: zHBM, a concept model stacking HBM directly above AI accelerator chips; zNAND-O, an edge-AI NAND design in four- and eight-layer versions; and V10 BV-NAND, a fully specified 400-plus-layer chip built with wafer bonding that raises density about 58 percent over the prior V9 generation.

Can enterprise buyers order zHBM or zNAND-O today?

No. Samsung showed zHBM and zNAND-O as concept models alongside products it already ships, including HBM4E samples, an HBM5 preview, and existing enterprise drives. Only V10 BV-NAND was presented as a fully specified chip, and independent analysis still flags bonding-yield, die-testing, and thermal-extraction hurdles before any of the three ship at data center scale.

Does this roadmap ease the current AI memory shortage?

Not in 2026. Samsung already holds about 34 percent of data center memory revenue and HBM is projected to consume around a quarter of all DRAM wafer output this year, with the data center off-chip memory market estimated to grow from about $17.1 billion in 2025 to $96.8 billion in 2026. A concept unveiling adds no supply to that market this year.

What should a European or UK buyer do with this news?

Treat the current memory shortage as the budget baseline for the next one to two cycles rather than a temporary condition. Lock in supply and pricing commitments now, while the constraint is known, instead of delaying a contract on the assumption that zHBM-class density gains will reach shipping servers within this year's or next year's refresh.

Samsung's vision models name a real bottleneck honestly, but a slide deck does not ship a wafer, and any owner budgeting AI infrastructure into 2027 should treat this roadmap as a reason to secure today's supply, not a reason to wait for tomorrow's.

SamsungMemoryAI InfrastructureFMS 2026Data Center

A 14 Percent Crash Will Not Cut Your Memory Bill

Samsung fell 13.4 percent and SK Hynix 14.7 percent on a Chinese lithography tool. That tool makes commodity DRAM, not HBM. What it changes in your memory contract, and when.

6 min read

Consumer Tech

The Memory Shock That Hit Xbox Just Hit Samsung

Samsung's Galaxy Z Fold8 hit 2,599 euros days after Microsoft raised Xbox prices for the same reason: a DRAM shock hitting unrelated device categories.

4 min read

Decisions

The Codename Worked. The Trading Controls Did Not

An indictment unsealed on 24 July says two Volkswagen engineers traded on a joint venture known internally as Project Climb. The control that failed is a common one.

5 min read

← All articles