How to read this page. Every maker in our Memory section uses the same 16 sections in the same order, the same shape as our AI Chips brand guides. Each section starts Simple, then goes Deeper, then Expert: stop wherever you have what you need. The small numbers are sources: click one to open the original document. Where a company has not published something, we say so rather than estimate.
1.At a glance
SimpleStart here
Samsung is the world's largest memory maker. It makes both kinds of memory that matter: DRAM, the fast working memory a computer forgets when you switch it off, and NAND flash, the slower storage that remembers 1. It also makes HBM, the stacked DRAM that sits beside AI accelerators 2.
What it makesDRAM, NAND flash and HBM, plus the SSDs and modules built from them
1
How bigKRW 120.8 trillion of memory revenue in one quarter, Q2 2026
3
Who makes itSamsung's own fabs, mostly in Korea plus Xi'an in China
4
Can you buy direct?Not the chips. Samsung publishes no price for any memory chip
5
DeeperThe detail
Samsung reports memory inside a division it calls DS, for Device Solutions. In Q2 2026 the DS Division posted "KRW 127.5 trillion in consolidated revenue and KRW 89.2 trillion in operating profit", both company records 6, of which the memory line alone was KRW 120.8 trillion of revenue 3. Samsung does not break out an operating profit for memory on its own — that figure exists only at DS-Division level 3.
ExpertFor specialists
Samsung's own investor deck puts its 2025 share at "DRAM 34% NAND 33%" on $74 billion of memory revenue, though the share figures are Samsung citing IDC rather than measuring itself 1. Its distinctive claim is scope rather than size: at GTC 2026 it called itself "the industry's only semiconductor company offering a total AI solution spanning memory, logic, foundry and advanced packaging" 7.
2.Company card
SimpleStart here
Samsung's memory business is one division of Samsung Electronics, a company that also makes phones and televisions. It is not a separate company.
Samsung's own filings say "listed its shares on the Korea Stock Exchange in 1975" without giving a ticker code 9. The widely-quoted code 005930 does not appear in the Samsung documents checked here, so this page does not state one.
DeeperThe detail
Samsung's November 2025 leadership release names Young Hyun Jun as "Vice Chairman and Head of Device Solutions (DS) Division", appointed "as co-CEOs of the Company", and states that "Vice Chairman Jun will remain as Head of Memory Business" 11 — one person holding both the division and the memory business. Samsung's board page gives his title as "Vice Chairman & CEO (DS Division)" and spells the name "YoungHyun Jun", one word 10; this page uses the spelling from the newsroom release.
ExpertFor specialists
Below him, the people Samsung quotes on memory in its own releases are Sang Joon Hwang, "Executive Vice President and Head of Memory Development" 2, Jin-Yub Lee, "Executive Vice President and Head of Flash Product & Technology" 12, and Jooyoung Lee, "Executive Vice President of DRAM Product & Technology" 13. Samsung publishes no separate memory-division headcount, no memory-only balance sheet and no memory-only operating profit 3.
3.The story
SimpleStart here
- 1974 Samsung Electronics acquires Hankook Semiconductor — the start of the chip business 14.
- 1983 Develops a 64Kb DRAM and establishes the Giheung campus 14.
- 1992 Two milestones in one year on Samsung's timeline: "Develops industry's first 64Mb DRAM" and "Achieves world's largest DRAM market share" 14.
- 2013 First mass production of 3D V-NAND, the stacked flash design still used today 14.
- 2017 Begins production at Pyeongtaek 14 15.
- 2026 Begins mass production of HBM4 and ships it commercially 2.
DeeperThe detail
Samsung dates its own leadership from the early 1990s: "We have been the leading memory provider in the industry for the past 30 years", and it marks 1993 with two entries, "Opens industry's first 200mm fabrication line (Line 5)" and "Achieves world's largest memory market share" 14. Its own framing of the 1974 start is that "we took our first step into uncharted territory" 16.
ExpertFor specialists
The published timeline is worth reading for what it leaves out. It runs from 1974 to 2024 and records LPDDR4, DDR5, CXL, LPCAMM2, GDDR7 and every V-NAND generation — and contains no HBM entry of any kind: no HBM2, HBM2E, HBM3, HBM3E or HBM4 14. For the product line that now dominates coverage of Samsung's memory business, its own official history is silent.
4.The lineup
DeeperThe detail
Samsung also sells module formats built for AI servers. SOCAMM2 is an LPDDR-based server module Samsung says "delivers more than twice the bandwidth of traditional RDIMM while consuming over 55% less power" 24. Its status is stated two ways in Samsung's own material: the GTC release says it "is currently in mass production, an industry-first to reach that milestone" 7, while the product blog describes "customer samples already being supplied" 24. The Q1 2026 results release settles it, recording "mass product sales of HBM4 and SOCAMM2 for NVIDIA's Vera Rubin platform" 25.
ExpertFor specialists
Samsung's investor deck lays out the generational ladders plainly: V-NAND as "V4 V5 V6 V7 V8 V9 V10 V11 Next" and DRAM as "28nm 25nm 20nm 1xnm 1ynm 1znm" then "1anm 1bnm 1cnm" 1. Beyond those, V10 BV-NAND is announced with "more than 400 layers" and, per Samsung's technical post, "a more than 58% increase in bit density over the previous generation" 12 26 — but neither Samsung page says V10 is sampling or in production 26. zHBM and HBM5 appear only as concepts 12, and this page treats them as such.
5.How it’s made
SimpleStart here
Samsung makes its memory in its own fabs, on processes it names by generation rather than by a plain nanometre figure.
- DRAM: Samsung's current leading node is what it calls "6th-generation 10 nanometer (nm)-class DRAM process (1c)" 2.
- NAND: the current shipping generation is 9th-gen V-NAND, in mass production since April 2024 21.
- Packaging: HBM stacks are built and tested by Samsung itself, alongside its logic and foundry work 7.
DeeperThe detail
The naming needs care, because Samsung uses two systems and never publishes a map between them. Its HBM4 release says "6th-generation 10 nanometer (nm)-class DRAM process (1c)" 2, while its DDR5 release calls the same era of technology "12 nanometer (nm)-class" throughout 13. Both are Samsung's own words; they are not stated to be the same thing, and this page does not assert that they are.
ExpertFor specialists
Samsung has announced no standalone release for 1c development or 1c mass production. The node appears only inside the HBM4 announcement 2 and in the Q3 2025 results line "the Company plans to proactively respond with capacity expansion in 1c" 18. Looking forward, Samsung and ASML said in September 2026 that Samsung plans to bring High-NA EUV into DRAM production by 2028, which Samsung describes as "the first time in the industry" for high-volume memory 27. Samsung does not publish the layer count of its 9th-gen V-NAND; its naming is generational only 21.
6.Where it’s made: fabs and addresses
SimpleStart here
Samsung publishes street addresses for its chip sites, which is unusual in this industry. The memory fabs are in Korea, with one in China.
DeeperThe detail
Pyeongtaek opened in July 2017 making "fourth-generation V-NAND (64 layers)", which Samsung then called "currently the largest single Fab in the industry", built with a planned KRW 30 trillion of investment through 2021 15. Xi'an opened in 2014 to make "Samsung's advanced NAND flash memory chips: 3D V-NAND" 28. For Xi'an, that 2014 release is effectively the last word: Samsung publishes no investment figure, no capacity and no current process node for the site 28 4.
ExpertFor specialists
The only capacity figure Samsung publishes is an aggregate. Its 1Q 2026 filing gives memory output as "658,532,038 (1,000 units)" in 1Gb equivalents at a utilisation rate of 100%, where "The utilization rate was calculated as the actual production time relative to the potential production capacity" 9. That is the whole memory business in one number: not per fab, not in wafers, and not split between DRAM and NAND. Anyone quoting a per-site Samsung memory capacity is not quoting Samsung 8.
7.How the memory works
SimpleStart here
The two products work on opposite principles. DRAM stores each bit as a charge in a tiny capacitor, which leaks — so the chip rewrites every bit thousands of times a second, and loses everything when power goes. NAND traps charge in a cell that holds it without power, which is why your files survive a reboot.
Samsung's NAND is V-NAND: instead of shrinking cells sideways, it stacks them vertically, which is how capacity keeps rising 14.
DeeperThe detail
HBM is a third arrangement rather than a third technology. It is ordinary DRAM dies stacked on top of each other and wired vertically, sitting right next to the processor instead of across the board. The point is width, not raw speed per wire: Samsung's HBM4 delivers "a maximum of 3.3 terabytes-per-second (TB/s)" per stack 2, where DDR5 runs at up to 8,000 megabits per second on each individual pin 5 — about 1 gigabyte per second per pin. HBM wins by having thousands of those connections at once rather than by making any one of them dramatically faster.
ExpertFor specialists
Samsung's HBM4 splits the job across two different processes: the memory dies use the "6th-generation 10 nanometer (nm)-class DRAM process (1c)", while the base die underneath — the logic layer that talks to the processor — is built on a "4nm logic process" 2. That combination is the practical argument behind Samsung's "memory, logic, foundry and advanced packaging" claim 7: an HBM stack is now a memory product that needs a logic fab to build.
8.Specs explained
DeeperThe detail
Two of Samsung's own pages disagree about HBM3E, and this page will not blend them. The February 2024 development release put the 36GB part at "up to 1,280 gigabytes per second (GB/s)" 30; the current HBM3E product page gives "bandwidth of up to 1,180 gigabytes per second (GB/s)" for the same 36GB capacity 29. Samsung has not published a correction reconciling the two, so both are reported here with their dates.
ExpertFor specialists
Samsung's HBM4 is deliberately faster than the standard it implements. JEDEC's HBM4 specification, JESD270-4, sets 8 Gb/s per pin across 32 channels for up to 2 TB/s per stack and a 64GB maximum 31; Samsung's part runs at 11.7 Gbps and it says so directly, describing it as "exceeding the industry standard of 8Gbps by approximately 46%" 2. The same pattern holds for LPDDR6: JEDEC's JESD209-6 defines the architecture — dual sub-channels, on-the-fly burst length, Per Row Activation Counting — without stating a maximum data rate 32, so Samsung's 10.7 Gbps is a product claim, not a ceiling set by the standard 19. Note also that Samsung's DDR5 product page is internally inconsistent, giving both 7,200 Mbps and 8,000 Mbps on the same page 5.
9.Official pricing
SimpleStart here
Samsung publishes no price for any memory chip. Its DRAM, HBM, GDDR and NAND pages carry specifications and a contact form, and no figure of any kind 5 29 20. Chips are sold through negotiated contracts and distributors, not a price list. It does publish how much its average selling price has moved, in its filings — see Expert below 9.
Finished Samsung-branded drives are different — those do have published prices 22.
DeeperThe detail
Those first two rows are worth reading together. A 4TB 9100 PRO launched at $549.99 in March 2025; by the time of this check a 2TB 990 EVO Plus was listed at $579.99 — half the capacity, a higher price, eighteen months apart 22 33.
ExpertFor specialists
Samsung does not publish price levels, but it does publish price movements — and its filings contain two different memory figures that are easy to mix up, because they run in opposite directions.
Under the heading "Changes in average selling price", Samsung's 1Q 2026 report states "Memory saw an increase of approximately 146%" 9. That is Samsung's own selling price — what its memory business charges. Separately, under "Production material price trends", the same report notes "the price of memory for mobile increased by around 107%" 9, and the half-year report gives "prices increased by approximately 6% for mobile APs—a major production material for the DX Division—and 211% for memory for mobile" 8. Those are what Samsung's own phone division pays to buy memory as an input.
So Samsung is simultaneously the seller enjoying a 146% ASP rise and a buyer absorbing a 211% cost rise on mobile memory, and it discloses both in the same document 9 8. Neither is a price list, but together they are an official, dated measure of the 2026 memory surge from both sides of it.
10.The HBM position
SimpleStart here
HBM is the memory that sits beside AI chips, and it is the most contested part of Samsung's business. The plain position as of this check: Samsung's HBM4 is in mass production and has shipped commercially 2, and Samsung itself names NVIDIA as the customer 25.
DeeperThe detail
Samsung's February 2026 release is unambiguous: it "has begun mass production of its industry-leading HBM4 and has shipped commercial products to customers" 2. The customer is named in the Q1 2026 results, which record "the industry's first mass product sales of HBM4 and SOCAMM2 for NVIDIA's Vera Rubin platform" 25, and the GTC release describes the part as "designed for the NVIDIA Vera Rubin platform" 7. HBM4E samples followed in May 2026 17, confirmed again in the Q2 results as "the industry's first HBM4E samples to major customers" 6.
ExpertFor specialists
What Samsung has not said matters as much, because a great deal of coverage fills these gaps. Samsung has published no statement that it passed NVIDIA's HBM3E 12-high qualification — its own wording stays at HBM3E being "in mass production and being sold to all related customers", with no customer named 18. It names no HBM3E customer at all. It gives no mass-production date for HBM4E, only that production will be "aligned with customer schedules" 17. And it publishes no claim that its 2026 HBM output is sold out. Each of those is widely reported; none is sourced to Samsung here, so none is stated as fact on this page.
11.Who buys it and why
SimpleStart here
Almost everything with a screen or a server rack in it contains memory from one of three companies, and Samsung is the largest. Its customers are the people building the finished thing: phone makers, PC makers, cloud operators and, now, AI accelerator makers 1.
DeeperThe detail
The one buyer Samsung names in its own memory material is NVIDIA, for HBM4 and SOCAMM2 on the Vera Rubin platform 25 7. Beyond that, Samsung describes demand by category rather than by customer: in July 2026 it said it expects "robust demand centered on servers stemming from continued AI infrastructure capex and broader adoption of agentic AI" with growth in "server DRAM, eSSDs, and HBM" accelerating 6.
ExpertFor specialists
Samsung's stated commercial posture in 2026 is scarcity management rather than volume growth: the Memory Business "achieved another record-breaking quarter by proactively addressing AI demand despite limited capacity with a primary focus on server products" 6. Its plan is to "lead the market by addressing strong demand with a focus on high-value-added products, such as HBM4, DDR5, SOCAMM2" 6, and it expects the market to stay "undersupplied, despite partial demand moderation in mobile and PCs" 6. In an undersupplied market, who gets allocation is the commercial decision, and Samsung does not publish how it makes it.
12.Hidden in plain sight
SimpleStart here
Samsung makes no memory in Texas. The Taylor, Texas project, widely described in coverage of Samsung's US expansion, is a logic and foundry site: Samsung's own page describes "two new leading-edge logic fabs and research and development (R&D) in Taylor" 34.
DeeperThe detail
Samsung's CHIPS Act factsheet sets out the whole US plan: Taylor gets "two fab projects, a R&D facility, and an advanced packaging facility" at nodes of "4nm + 2nm", and Austin gets "expansion and modernization of the existing semiconductor manufacturing campus" 35. Samsung's investment there is real and large — its Taylor page gives "$17 BILLION" as "Samsung's initial minimum investment in the Taylor fab", against "up to $4.745 billion in direct funding under the CHIPS and Science Act" 34 — but none of it is described by Samsung as memory production. Note that the April 2024 factsheet quotes the earlier, preliminary figure of "$6.4 billion in direct funding from the U.S. Department of Commerce" 35; the $4.745 billion on Samsung's current page is the finalised award, and this page leads with that.
ExpertFor specialists
The second thing hidden in plain sight is the timeline gap noted earlier: Samsung's official milestone history records every V-NAND generation and every DRAM first from 64Kb in 1983 onward, and does not mention HBM once 14. A third: Samsung's current Taylor page gives the "Year when all CHIPS and Science Act facilities will be operational" as 2030, against a groundbreaking year of 2022 34 — an eight-year build. Earlier and much nearer dates circulated widely when the project was announced; this page quotes only the figure Samsung publishes now.
13.Weak spots
SimpleStart here
- Capacity, in its own words. Samsung says it met Q2 2026 demand "despite limited capacity" 6.
- No HBM history. Its official timeline has no HBM milestone at all, in the generation where HBM decides the league table 14.
- Unclear own numbers. Its DDR5 page gives two different top speeds, and its HBM3E bandwidth differs between two Samsung pages 5 30 29.
DeeperThe detail
The disclosure gap is a real weakness for anyone trying to check Samsung's position. There is no Samsung release announcing 1c development or production 2; no published layer count for 9th-gen V-NAND 21; no Samsung statement of GDDR7 mass production, only a product page listing the speed 20 36; and no stated production status for LPDDR6 19. In each case the widely-cited figure comes from somewhere other than Samsung.
ExpertFor specialists
Structurally, the weak spot is that memory profit is not separable from the rest of DS. Samsung reports DS-Division operating profit only 3, so the margin on HBM specifically — the number that would settle most arguments about whether Samsung is winning or losing the AI memory race — is not published by Samsung at all, and cannot be derived from its filings. The second is China: Xi'an is a significant NAND site whose current technology Samsung has not described publicly since 2014 28.
14.India angle
SimpleStart here
Samsung has no memory fab in India. What it has is research: Samsung Semiconductor India Research (SSIR), in Bengaluru, Karnataka 37.
DeeperThe detail
Samsung describes SSIR as "part of the global network of Samsung Electronics Co., Ltd. for providing component solutions, featuring industry-leading technologies in the areas of System LSI, Memory and Foundry in Asia Pacific region" 37. Its memory scope is given as "DRAM and Flash technologies for various applications including servers, big data, AI/HPC, PC, IoT, 5G, automotive and more", covering DRAM, SSD, eStorage and MCP work 37. Samsung publishes neither an establishment date nor a headcount for it 37.
ExpertFor specialists
Samsung does publish Indian prices for finished drives — the P9 and P7 portable SSDs were launched with rupee pricing from ₹27,839 to ₹259,199 23 — so India is a priced retail market for Samsung storage even though it is not a manufacturing location for Samsung memory. No Samsung source found in this check announces memory manufacturing, assembly or test investment in India, and this page states that as a finding rather than leaving it unsaid 37.
15.What’s next
SimpleStart here
Only what Samsung has actually announced:
- HBM4E: samples shipped; mass production "aligned with customer schedules", no date 17.
- V10 BV-NAND: announced with over 400 layers; no production status stated 12 26.
- High-NA EUV in DRAM: planned by 2028, with ASML 27.
DeeperThe detail
Samsung's own H2 2026 outlook expects demand "centered on servers", with "growth in demand for server DRAM, eSSDs, and HBM" accelerating and the market staying undersupplied 6. Its stated response is product mix rather than volume: HBM4, DDR5 and SOCAMM2 6.
ExpertFor specialists
Further out, Samsung's FMS 2026 material describes zHBM as a concept promising "approximately eight times the performance of HBM5" and "more than 10 times the memory density of HBM5" 12. Both zHBM and HBM5 are concepts in Samsung's own framing, with no product claim attached, and this page reports them that way. The nearer and more consequential item is the ASML agreement: Young Hyun Jun's framing of it — "The AI era is transforming the semiconductor industry and increasing the importance of technological innovation across the entire value chain" — sits alongside a concrete 2028 target for High-NA EUV in high-volume DRAM 27.