SK HynixPrefer Sterling in Google
Add sterlingcharts.com as a preferred source so our charts can appear with a preferred badge in Google Top Stories, AI Mode, and AI Overviews.
Register for a premium account to gain access to Sterling AI.
Get StartedThings you can ask Sterling:
| Period | DRAM ASP QoQ | NAND ASP QoQ |
|---|---|---|
| Q2 2025 | Up low single-digit % | Down high single-digit % |
| Q3 2025 | Up mid-single-digit % | Up low-teen % |
| Q4 2025 | Up mid-20% | Up low-30% |
| Q1 2026 | Up mid-60% | Up mid-70% |
| Q2 2026 | Up approximately 30% | Up mid-50% |
| Period | DRAM bits QoQ | NAND bits QoQ | Actual or plan |
|---|---|---|---|
| Q2 2025 | Mid-20% | Over 70% | Actual (call) |
| Q3 2025 | High single-digit % | Down mid-single-digit % | Actual (call) |
| Q4 2025 | Low single-digit % | Approximately 10% | Actual (call) |
| Q1 2026 | Similar to Q4 2025 (flat vs prior quarter) | Down approximately 10% | Actual (call) |
| Q2 2026 | High single-digit % | Mid-teen % | Actual (call) |
| Q3 2026 | Approximately 10% | Low single-digit % | Company plan from the Q2 2026 call |
| Period | DRAM share of revenue (%) | NAND share of revenue (%) |
|---|---|---|
| Q2 2025 | 77 | Not printed as a standalone NAND % on the recap used here |
| Q1 2026 | 78 | 21 |
| Q2 2026 | 73 | 27 |
| Period | Printed product facts | Customer / status (company wording) |
|---|---|---|
| HBM3 (2022) | Company later described itself as the industry's first provider to mass-produce HBM3 in 2022 | Not a current-generation volume story. Kept as the company's own first-mover claim on the HBM4 sample release |
| HBM3E 8-layer (Mar 2024) | Delivered to customers in March 2024, described as first in the industry | Customers not named in the September 2024 newsroom note |
| HBM3E 12-layer (Sep 2024) | Volume production began 26 Sep 2024. 36 GB stack. DRAM chips 40% thinner than the 8-layer product. Operating speed 9.6 Gbps. Advanced MR-MUF | Company said it would supply mass-produced 12-layer product to customers within 2024. Customers not named |
| HBM4 12-layer samples | Company shipped 12-layer HBM4 samples. Bandwidth more than 2 TB/s. Capacity 36 GB. 2,048 I/O. Mass-production target at sample time was 2H 2025 | Samples to major customers, not named in the newsroom note |
| Q2 2026 HBM4 mass shipments | Mass production shipments of HBM4 began in Q2 2026. Full ramp planned in 2H 2026. Call: yield and quality nearing mature HBM3E levels | Q2 2026 call named NVIDIA as a major AI customer in the HBM collaboration discussion. Other key customers not named |
| HBM4E (H1 2026 samples) | Sample shipments completed in the first half of 2026. Company said it used a mature, mass-production-stable process node | Supplied to a major customer in 1H 2026. Customer not named |
| Period | Printed HBM comment | What was not disclosed |
|---|---|---|
| Q1 2025 | HBM demand expected to approximately double vs 2024. 12-layer HBM3E expected to account for over 50% of total HBM3E revenues in Q2 2025 | Not HBM as a percent of DRAM. Not an HBM dollar figure |
| FY 2025 | HBM revenue more than doubled year-on-year | No HBM dollar amount and no HBM percent of DRAM for the year or for Q4 |
| Q2 2026 | Began mass shipments of HBM4. HBM4E sample shipments completed in H1 2026. Call: HBM4 yield and quality nearing HBM3E maturity | No HBM percent of DRAM. No HBM revenue dollar figure for the quarter |
| Period | 321-layer NAND comment |
|---|---|
| Q1 2026 | Q2 2026 call: in the previous quarter, 321-layer accounted for the largest proportion of NAND production |
| Q2 2026 | Newsroom: 321-layer products already represent the largest share of total production |
| YE 2026 plan | Approximately 50% of domestic production capacity. Capacity target, not a current mix % |
The HBM kingpin — supplies the memory that makes AI GPUs actually work.
HBM3E
High Bandwidth MemoryPrimary supplier to NVIDIA for H100/B200
HBM4
Next-Gen HBMIn development for 2026 AI accelerators
DDR5
Server DRAMHigh-performance server memory
NAND Flash
StorageEnterprise SSD and data center storage
Market Share
~50% of HBM market (dominant), ~28% of total DRAM
Competitive Moat
First-mover in HBM3E, exclusive NVIDIA supply relationship, process technology lead
Key Risk
Revenue highly concentrated in a cyclical memory market; Samsung and Micron catching up
Without SK Hynix's HBM chips, NVIDIA's GPUs can't function. It's that simple. SK Hynix controls half the HBM market and is NVIDIA's primary memory supplier. They're the hidden bottleneck of the entire AI buildout.
Founded as Hyundai Electronics in Icheon, South Korea by Hyundai Group founder Chung Ju-yung; initial focus on DRAM manufacturing as Korea's second memory player behind Samsung.
Renamed Hyundai Electronics Industries; consumer electronics arm and memory unit operated separately under the Hyundai chaebol umbrella through the 1997 Asian Financial Crisis.
Forced merger with rival LG Semicon under Korean government Big Deal industrial policy; combined entity briefly became world’s largest DRAM maker by capacity.
Hyundai Electronics renamed Hynix Semiconductor March 26 amid debt restructuring; Korean creditor banks took control after the 1997 financial crisis nearly bankrupted the company.
Acquired by SK Group for B February 14; renamed SK Hynix and entered a 14-year run of stable strategic capex under Chairman Chey Tae-won.
Released the world's first HBM (High Bandwidth Memory) device, co-developed with AMD on TSV (through-silicon-via) packaging; foundation of all modern AI accelerators a decade later.
Joined Bain-led consortium taking 49% stake in Toshiba Memory (now Kioxia) for ~B; secured non-controlling NAND access to balance Samsung-dominated supply chain.
Announced B M16 chip-cluster investment in Yongin, Korea; the cluster ultimately housed leading-edge DRAM and HBM production lines coming online through 2027.
Announced B acquisition of Intel’s NAND/SSD business October 19; closed in two phases through 2025 with Solidigm as standalone subsidiary running ex-Intel Dalian fab.
Began industry-first HBM3 mass production in June, sole supplier to NVIDIA H100 at launch; locked in 5+ years of dominant AI memory share and became NVIDIA's most strategic supplier.
Granted one-year US license waiver for advanced equipment imports into its Wuxi DRAM fab, easing immediate concerns about strandable Chinese memory capacity.
Received indefinite VEU authorization from US Commerce, securing Wuxi DRAM fab equipment access long-term; protected ~40% of SK Hynix DRAM capacity that resides in China.
Began industry-first HBM3e 8-layer mass production for NVIDIA H200 ramp; SK Hynix supplied roughly 80% of NVIDIA's HBM3e through 2024 alongside Micron's 24GB units.
Began HBM3e 12-layer mass production at 36GB/9.6Gbps in September, six months after the 8-layer ramp; locked in 12-layer supply for Blackwell B200 and B300.
December 2 BIS rules added country-wide HBM controls, restricting any HBM exports to China above 2 GB/s/mm2; effectively cut SK Hynix off from Chinese AI accelerator customers.
Delivered industry-first 12-layer HBM4 samples to NVIDIA, ahead of plan and gating Rubin platform launch; cemented projected ~70% share of Rubin HBM4 demand.
Delivered final paid HBM4 samples to NVIDIA targeting February 2026 mass production for Rubin; secured estimated 70% of NVIDIA HBM4 supply allocation.
Began HBM4 mass production for NVIDIA Rubin; shipped 12-layer stacks at 9.6Gbps with 2TB/s bandwidth per stack, six months ahead of original 2026 H2 target.