SK Hynix
Last reviewed
May 25, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v1 ยท 4,065 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 25, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v1 ยท 4,065 words
Add missing citations, update stale details, or suggest a clearer explanation.
SK Hynix Inc. (KRX:000660) is a South Korean memory semiconductor manufacturer headquartered in Icheon, Gyeonggi-do, and the world's largest producer of high bandwidth memory (HBM) used in AI accelerators.[1][2] In the first quarter of 2025, SK Hynix overtook Samsung Electronics to become the largest DRAM supplier by revenue for the first time in its history, capturing about 36% of the global DRAM market on the strength of HBM sales.[3] The company was the first in the industry to mass-produce hbm3e in March 2024 and the first to ship 12-layer HBM3E with 36 GB capacity in September 2024, supplying chips to NVIDIA, AMD, and other AI accelerator vendors.[4][5] In March 2025 SK Hynix delivered the world's first 12-layer hbm4 samples to customers, and in fiscal first quarter 2026 it posted record revenue of 52.58 trillion won (about US$36 billion) and an operating margin of 72%, the highest in its history.[6][7] The company is part of the SK Group, the third-largest chaebol in South Korea, and its wholly owned U.S. subsidiary Solidigm operates the NAND flash and SSD business acquired from Intel between 2021 and 2025 for US$9 billion.[8][9]
SK Hynix traces its origins to Hyundai Electronics Industries Co., Ltd., founded on 13 February 1983 by Chung Ju-yung, the founder of the Hyundai conglomerate, who positioned the new affiliate to participate in the rapidly expanding global semiconductor industry.[2][10] Hyundai Electronics shipped its first 16 Kb static RAM in 1984 and entered DRAM production in 1985, growing through the late 1980s and early 1990s on the back of Korea's broader memory chip expansion against Japanese and U.S. competitors.[10] By 1992 Hyundai had become the world's ninth largest DRAM manufacturer, and by the mid-1990s the company ranked among the top 20 global semiconductor producers.[2][10]
In the wake of the 1997 Asian financial crisis, the South Korean government pressed the country's chaebols to consolidate overlapping business lines under a program known as the "big deal" restructuring.[2] In May 1999 Hyundai Electronics acquired LG Semicon from the LG Group for about US$2.1 billion, creating a single Korean memory rival to Samsung Electronics and assuming LG's debts in the process.[2][11] The combined company controlled more than 200,000 wafers per month of capacity, comparable to Micron Technology and ahead of most non-Samsung competitors.[11]
In March 2001 Hyundai Electronics rebranded as Hynix Semiconductor Inc., a contraction of "high" and "electronics," and began spinning off non-memory units including Hyundai Curitel and Hyundai SysComm to focus on DRAM and NAND.[2][10] The 2001 chip downturn, in which global memory prices fell by roughly 80%, forced the company into a creditor-led workout that left a consortium of South Korean banks owning a controlling stake.[2] Hynix operated under creditor control for the next decade, restoring profitability through the mid-2000s while continuing to invest in advanced fabs at Icheon and Cheongju.[10] In 2010 Hynix and AMD jointly proposed the High Bandwidth Memory specification to JEDEC, beginning a memory standards effort that would later define the company's competitive position in AI.[12]
On 14 February 2012 SK Telecom completed the purchase of a 21.05% stake in Hynix Semiconductor from creditors for about 3.4 trillion won (approximately US$3 billion), giving the SK Group a controlling interest and management rights.[13] The company held an inauguration ceremony on 26 March 2012 under the new name SK Hynix Inc., with the SK chairman Chey Tae-won committing to large capital expenditures to keep the company competitive against Samsung in DRAM and against Toshiba (later Kioxia) and Micron in NAND.[13] Under SK ownership SK Hynix continued to expand fabs in Wuxi, China, built the M14 fab in Icheon (operational from 2015), and pursued joint development with AMD on the first HBM generation, which entered production in 2014 on 29 nm process technology at Icheon.[12][14] In 2020 SK Hynix agreed to acquire Intel's NAND flash and solid state drive business for US$9 billion, the largest acquisition in South Korean semiconductor history at the time.[8][15]
SK Hynix sells three main product families: DRAM, NAND flash and enterprise SSDs (through the Solidigm subsidiary), and HBM stacks built from custom DRAM dies. The company's overall product strategy increasingly emphasizes high-margin AI memory, which in 2024 and 2025 accounted for an outsized share of revenue and profit.[1][16]
SK Hynix produces DDR5, LPDDR5X, GDDR6, GDDR7 and various server, mobile and graphics DRAM products on advanced sub-15 nm process nodes branded as 1a (about 14 nm), 1b (about 12 nm) and 1c (about 11 nm).[1][17] In October 2020 SK Hynix announced the world's first DDR5 DRAM product, an 8 Gb chip running at 4,800 Mbps, complying with the JEDEC DDR5 standard published earlier that year.[17] The company has launched a series of high-capacity server DDR5 modules including 128 GB and 256 GB RDIMMs used in cloud and AI training servers.[16] LPDDR5X mobile and server DRAM products from SK Hynix are deployed in Apple iPhones and in low-power compute attached memory modules (SOCAMM) used by NVIDIA in next-generation Grace and GB300 systems.[18]
HBM is a class of DRAM in which multiple memory dies are stacked vertically and connected through silicon vias (TSVs), then placed alongside a GPU or accelerator on a silicon interposer to deliver bandwidths of more than 1 TB/s per stack.[12] SK Hynix is the largest HBM supplier by both revenue and volume, with TrendForce and Counterpoint Research estimating the company's HBM market share at between 57% and 62% in 2025 quarters, ahead of Samsung Electronics and Micron Technology.[19] Customers for SK Hynix HBM include NVIDIA (H100, H200, B100, B200, GB200, GB300, and Vera Rubin generation accelerators), AMD Instinct MI300X and successors, Broadcom AI ASIC programs, and Intel Gaudi.[4][5][20]
SK Hynix produces 3D NAND flash for client SSDs, mobile UFS storage and embedded products, and operates the higher-density enterprise SSD line through Solidigm, the wholly owned subsidiary based in San Jose, California that absorbed Intel's NAND and SSD business.[8][9] In January 2026 SK Hynix announced that NAND set its highest annual revenue on record in fiscal 2025 and that 321-layer QLC NAND product development was complete, an industry-leading layer count for triple level cell and quad level cell 3D NAND.[6]
The High Bandwidth Memory standard was first proposed jointly by AMD and Hynix to JEDEC in 2010 and ratified as JESD235 in October 2013.[12] HBM was designed to overcome the bandwidth ceiling of GDDR by stacking DRAM dies in 3D using TSVs and connecting them to the host processor through a wide 1,024-bit interface on a 2.5D silicon interposer.[12] SK Hynix completed development of the first HBM chip in 2013 and shipped it in volume in 2015 inside AMD's Radeon R9 Fury X (Fiji) GPU, the first commercial product to use HBM.[12][14] The collaboration with AMD established SK Hynix's early lead in the HBM packaging process, including the company's proprietary Mass Reflow Molded Underfill (MR-MUF) bonding technique that filled gaps between stacked dies with a thermally conductive epoxy resin in a single reflow step.[4]
Between 2018 and 2020 the HBM market remained a niche, used primarily in GPGPU products from NVIDIA Tesla and AMD Instinct lines and in a handful of FPGA and HPC products. NVIDIA's pivot to large language model training and inference, beginning with the A100 in 2020 and accelerating with the H100 in 2022, transformed HBM into a strategic AI memory component.[14] SK Hynix mass-produced HBM2E in mid-2020 and HBM3 in June 2022, the first HBM3 generation to ship in volume.[14] By the end of 2023 HBM3 sales accounted for a low double-digit percentage of SK Hynix's DRAM revenue, but by the fourth quarter of 2024 HBM had grown to over 40% of DRAM revenue and supported a sharp expansion of operating margin from prior cycles.[1][21]
In March 2024 SK Hynix became the first memory maker to mass-produce HBM3E (Extended), starting with 8-high stacks of 24 GB capacity supplied to NVIDIA for the H200 and B200 product lines.[4][22] In September 2024 SK Hynix began volume production of the world's first 12-layer HBM3E with 36 GB capacity per stack and 9.6 Gb/s I/O speed, increasing capacity by 50% relative to the 8-high product without increasing stack height.[5][23] In November 2024 at the SK AI Summit, CEO Kwak Noh-Jung announced 16-layer HBM3E samples with 48 GB capacity targeted at customer qualification.[24] In March 2025 SK Hynix shipped the industry's first 12-layer HBM4 samples, delivering bandwidth of more than 2 TB/s per stack (over 60% faster than HBM3E) and again building stacks at 36 GB capacity.[25][26] HBM4 from SK Hynix was selected as primary memory for NVIDIA's Vera Rubin AI accelerator family.[25] In the fiscal 2025 results announcement on 28 January 2026, the company stated that it had completed preparations for HBM4 mass production in September 2025 and that HBM4E samples would be supplied in the second half of 2026 with mass production targeted for 2027.[6][7]
| Generation | Stack height | Capacity | I/O speed | Bandwidth per stack | SK Hynix mass production | Lead customer reference |
|---|---|---|---|---|---|---|
| HBM1 | 4-Hi | 1 GB | 1.0 Gb/s | 128 GB/s | 2015 | AMD Radeon R9 Fury X (Fiji) |
| HBM2 | 8-Hi | 8 GB | 2.4 Gb/s | 307 GB/s | 2018 | NVIDIA Tesla V100 |
| HBM2E | 8-Hi | 16 GB | 3.6 Gb/s | 460 GB/s | 2020 | NVIDIA A100 |
| HBM3 | 12-Hi | 24 GB | 6.4 Gb/s | 819 GB/s | June 2022 | NVIDIA H100 |
| HBM3E (8-Hi) | 8-Hi | 24 GB | 9.2 Gb/s | 1.18 TB/s | March 2024 | NVIDIA H200, B200 |
| HBM3E (12-Hi) | 12-Hi | 36 GB | 9.6 Gb/s | 1.23 TB/s | September 2024 | NVIDIA GB300 |
| HBM4 (12-Hi) | 12-Hi | 36 GB | >7 Gb/s (2,048-bit) | >2 TB/s | 2H 2025 (planned) | NVIDIA Vera Rubin |
Sources for the table: SK Hynix HBM development newsroom and product press releases.[4][5][14][25]
SK Hynix concentrated its HBM front-end capacity at the Cheongju campus and on advanced packaging lines at Icheon. The M15X fab at Cheongju, originally announced in 2022 with an investment of over 20 trillion won, was repurposed as a primary HBM front-end facility to meet AI memory demand.[27] The company pulled forward the M15X start by four months, with first equipment moved in during 2025 and DRAM production targeted to support HBM4 ramp at the site.[28] When M15X is fully operational, SK Hynix's HBM wafer capacity is expected to rise by 20% to 30% above 2024 levels, with monthly wafer starts of 55,000 to 60,000.[27]
In 2019 the South Korean government and SK Hynix announced the Yongin Semiconductor Cluster project, a planned four-fab campus on a 4.15 million square meter site in Yongin, Gyeonggi-do, with an initial investment commitment of 120 trillion won.[29] In October 2024 the SK Hynix board approved a first phase of about 9.4 trillion won to build the first fab and supporting facilities at Yongin, and in February 2026 the company added a further 21.6 trillion won (about US$15 billion) for first-fab completion by the end of 2030, bringing the first-fab investment to roughly 31 trillion won.[30][31] In 2024 SK Group chairman Chey Tae-won announced an expanded long-term plan calling for about 128 trillion won by 2028 and total Yongin investment of roughly 600 trillion won (about US$410 billion) when the four fabs and the surrounding supplier ecosystem are factored in over a multi-decade horizon, although the latter figure is a planning aspiration rather than a committed capex.[31] The Yongin cluster includes a 50-supplier industrial complex intended to anchor the domestic semiconductor supply chain.[29]
SK Hynix also operates the M14 (Icheon, opened 2015), M15 (Cheongju, opened 2018) and M16 (Icheon, opened 2021) fabs, alongside the C2 and C2F DRAM lines in Wuxi, China.[32] The Wuxi C2F expansion was completed in April 2019 at a cost of 950 billion won and added one-story cleanroom space comparable to the original C2 fab; in 2025 SK Hynix completed conversion of the Wuxi fab to the 1a (about 14 nm) DRAM node, with about 90% of the Wuxi line operating on 1a by year end.[32][33] Capital expenditure jumped sharply in 2024 and 2025 to support the HBM build-out, with SK Hynix guiding investors that 2026 capex would be the highest in company history.[1][16]
SK Hynix's HBM revenue is concentrated among a small number of AI accelerator customers. NVIDIA is by far the largest HBM customer, with priority allocations for HBM3E 8-Hi (H200 and B200), HBM3E 12-Hi (GB300), and HBM4 (Vera Rubin).[4][25][26] AMD is the second-largest HBM customer through the MI300X, MI325X and MI355X lines.[20] Broadcom and other ASIC vendors building custom AI accelerators for hyperscalers, and tsmc which packages many of these accelerators using its CoWoS interposer process, also rely on SK Hynix HBM.[20] Intel uses SK Hynix HBM in Gaudi 3 accelerators for AI training.[20]
In May 2024 SK Hynix's CEO Kwak Noh-Jung stated that the company's HBM production for 2024 was sold out and that 2025 supply was almost fully allocated, a position the company reiterated through the 2025 financial year as HBM demand continued to outrun supply.[34][35] In March 2025 SK Hynix indicated it expected to finalize allocations of all 2026 HBM output by mid-2025.[35] As of fiscal first quarter 2026 the company said HBM4 demand from customers for the next three years was already in excess of its installed and planned capacity.[7]
This concentration creates significant revenue dependence on AI capital expenditure cycles at NVIDIA's principal customers, including the hyperscale cloud providers Microsoft, Meta, Google, Amazon and Oracle, as well as on NVIDIA's own ability to ship Blackwell and Rubin systems on its planned cadence.[36] In Q3 2025 HBM accounted for over 40% of SK Hynix DRAM revenue, and HBM unit average selling prices (ASPs) ran several multiples above commodity DDR5.[21]
On 20 October 2020 SK Hynix announced an agreement to acquire Intel's NAND flash memory business for US$9 billion in a two-phase transaction.[8] The deal covered Intel's NAND SSD business, its NAND component and wafer manufacturing operations, and the Dalian, China NAND fabrication facility, while Intel retained the separate Optane 3D XPoint business.[8] In the first closing on 30 December 2021, SK Hynix paid US$7 billion and took over the Dalian fab, the SSD operations and associated intellectual property, branding the new subsidiary Solidigm and basing it in San Jose, California.[9][15] The second and final closing on 27 March 2025 transferred remaining NAND wafer manufacturing IP and R&D employees in exchange for the final US$1.9 billion, with adjustments.[15][37]
Solidigm sells enterprise QLC and TLC NAND SSDs, including the D5-P5430 and D5-P5336 high-capacity QLC drives that have become reference products for AI training and inference storage tiers.[9] Through Solidigm and SK Hynix Korea NAND operations, the combined group ranks second in global NAND flash market share by revenue behind Samsung in recent quarters, ahead of Kioxia, Micron and Western Digital (now Sandisk).[37]
In fiscal 2024 SK Hynix reported revenue of 66.193 trillion won and operating profit of 23.467 trillion won, an operating margin of 35%, both all-time highs that exceeded the prior 2018 peak of the memory super-cycle.[1] In fiscal 2025, revenue rose 47% year over year to 97.147 trillion won and operating profit doubled to 47.206 trillion won, lifting operating margin to 49% and net profit to 42.948 trillion won.[6] The company's board approved an additional shareholder return of 1 trillion won in dividends for FY2025 (1,500 won per share) and a treasury share cancellation of 15.3 million shares worth about 12.2 trillion won.[6]
In fiscal first quarter 2026 (the calendar quarter ending March 2026) SK Hynix reported revenue of 52.576 trillion won, operating profit of 37.610 trillion won and net profit of 40.346 trillion won, with operating margin reaching 72%, all of which were all-time quarterly records.[7][38] CNBC reported that operating profit roughly doubled quarter over quarter and rose more than fivefold year over year, attributable to AI-driven HBM and high-capacity server DRAM demand combined with rising commodity DRAM prices.[38] In April 2026 SK Hynix said HBM4E sample shipments would begin in the second half of 2026, with HBM4E mass production scheduled for 2027.[7]
The Q1 2026 operating margin of 72% exceeded the operating margin reported by NVIDIA in the comparable quarter, a measure cited in Korean media as an indicator of how far HBM pricing had outrun memory production costs during the AI capex cycle.[39] By the start of fiscal 2026 SK Hynix's market capitalization had surpassed 400 trillion won, placing it among the most valuable Asian companies and the second most valuable South Korean company after Samsung Electronics.[39]
SK Hynix sits at the intersection of several geopolitical pressures affecting the global memory industry. The company's Wuxi DRAM fab in China accounts for a significant portion of total DRAM capacity, and U.S. export controls beginning in October 2022 restricted shipment of advanced semiconductor manufacturing equipment to Chinese-located fabs, prompting Validated End User and other licensing arrangements that allow SK Hynix and Samsung to continue operating their China fabs under U.S. Department of Commerce authorizations.[40] In December 2023 the U.S. Department of Commerce granted SK Hynix and Samsung indefinite Validated End User status, providing a regulatory framework for continued operation in China.[40]
Under the U.S. CHIPS and Science Act, the U.S. Department of Commerce in late 2024 finalized a direct funding award of up to US$458 million to SK Hynix to support construction of an advanced HBM packaging fabrication and R&D facility in West Lafayette, Indiana, with total project costs of approximately US$3.87 billion and a planned start of mass production in the second half of 2028.[41][42] The Indiana plant is the first U.S. HBM advanced packaging facility and is intended to reduce U.S. dependence on East Asian advanced packaging for AI accelerator supply chains.[41] SK Hynix broke ground on the West Lafayette site in April 2026.[42]
In Korea, SK Hynix benefits from the Korean government's K-Chips Act and from tax credits for memory R&D and capital expenditure under the Restriction of Special Taxation Act, and the company participates in the Yongin megacluster as the anchor tenant alongside Samsung's K-2 cluster in Pyeongtaek.[29] Korea-Japan relations affect SK Hynix's supply of photoresist, hydrogen fluoride and fluorinated polyimides, materials that Japan briefly restricted in 2019, prompting Korean substitution efforts.[43]