SK Hynix
The world's dominant HBM memory supplier — 60-70% of AI GPU HBM, sole supplier of HBM4 to NVIDIA, and the single most supply-constrained company in the AI chip stack outside of TSMC CoWoS.
1. Core Product / Service
SK Hynix (KOSPI: 000660) is a South Korean memory semiconductor company and the global leader in High Bandwidth Memory (HBM) — the stacked DRAM used inside every NVIDIA AI GPU (H100, H200, B100, B200, GB200), every Google TPU, and every major AI accelerator.
Key products:
- HBM3e — 12-high stacks at 9.6 Gbps/pin, 1.15 TB/s bandwidth per stack; currently in mass production for NVIDIA H200/B200
- HBM4 — 16-high stacks; world's first HBM4 developed and readied for mass production at M16 Icheon + M15X Cheongju fabs; mass production began February 2026 [3]. Uses TSMC 12nm base die. Bandwidth >2 TB/s per stack.
- Conventional DRAM — less AI-exposed but core business
SK Hynix supplies approximately 60-70% of NVIDIA's HBM in the HBM4 era, with ~70% of NVIDIA's HBM4 orders [1]. The remainder goes to Samsung and Micron.
2. Target Users & Pain Points
- NVIDIA — the primary customer for HBM3e and HBM4; SK Hynix is NVIDIA's preferred HBM supplier
- Google — TPU uses SK Hynix HBM
- AMD — MI300X/MI325X uses HBM3e
- Broadcom — for custom AI accelerators (Google TPU, Meta MTIA)
- Microsoft, Amazon — AI accelerators (Maia, Trainium) use some HBM
Pain solved: AI GPUs cannot function without HBM. Without SK Hynix's HBM, there is no AI accelerator. HBM4 is physically different from HBM3e and requires new packaging equipment (be-semiconductor TC bonders + hybrid bonders, asmt TCB) and new qualification cycles. SK Hynix's mass production readiness at M16/M15X fabs in Feb 2026 is the gate for NVIDIA's Rubin GPU production in 2H 2026.
3. Competitive Landscape
| Company | HBM Position | vs SK Hynix |
|---|---|---|
| samsung-hbm | #2; 12-Hi HBM3e struggled with yields; HBM4 qualification in progress | Samsung has <20% share vs SK Hynix's 60-70%; Samsung's HBM3e had failures with NVIDIA qualification in 2024 |
| Micron | #3; HBM3e recently qualified for some AI accelerators | Micron has ~10% share; focused more on mainstream inference |
| Conventional DRAM competitors | DRAM market shares differ | HBM is a premium segment with different customers |
SK Hynix's moat: manufacturing execution — proven HBM yield at 12-Hi and now 16-Hi, massive capital investment in M16/M15X fabs, and TSMC partnership for base die. The base-die partnership (TSMC 12nm for HBM4) is a significant competitive advantage that samsung-hbm and Micron don't have.
4. Unique Observations
- HBM4 mass production started Feb 12, 2026 at M16 Icheon + M15X Cheongju — world's first commercial HBM4 [3]. This is the pivotal event for 2H 2026 AI GPU supply.
- NVIDIA HBM4 orders: ~70% from SK Hynix [1]. This is near-monopoly supply for the most critical AI chip component.
- HBM3e: SK Hynix first to ship 36 GB in Q3 2023, 48 GB variant in Q4 2024 — consistently first-to-market.
- Market share estimates: SK Hynix ~60%, Samsung ~20%, Micron remainder (2026 outlook) [1].
- Pricing power: HBM price increased ~5x from 2023 to 2026. Even at these prices, demand exceeds supply — AI GPU makers would pay more if that would improve allocation priority.
- Wuxi fab capacity: SK Hynix expanded 1b nm DRAM production from 10,000 wpm to 90,000 wpm in mid-2024 — massive capacity investment [PatSnap].
- The key bottleneck: HBM4 is not just a SK Hynix product problem — it's also limited by be-semiconductor TC bonder availability and asmt TCB equipment lead times (12-18 months). Even if SK Hynix has the HBM4 chips ready, packaging them is limited by equipment supply.
- TSMC CoWoS bottleneck interaction: SK Hynix ships HBM to TSMC for CoWoS packaging. TSMC's CoWoS capacity is allocated (NVIDIA holds 60% of CoWoS [3]). Even with sufficient HBM4 supply, the combined TSMC CoWoS + HBM4 allocation is the true ceiling on Rubin GPU出货量 in 2026-2027.
5. Financials / Funding
- Listed: KOSPI: 000660; market cap ~$130B+ (2026)
- Revenue: HBM-dominant memory business; FY2024-FY2025 memory recovery driven by AI; SK Hynix posted record Q2 earnings (2025) riding AI wave [source: SemiWiki]
- Capital expenditure: massive — M16 Icheon expansion, M15X Cheongju expansion, Wuxi capacity
- HBM4 revenue: expected to command premium pricing vs HBM3e due to 16-Hi complexity
6. People & Relationships
- TSMC relationship: TSMC supplies base dice for HBM4 under a commercial agreement — this is a critical partnership for SK Hynix's HBM4 competitiveness
- NVIDIA relationship: preferred HBM4 supplier; close technical collaboration for HBM co-design
- Packaging equipment suppliers: be-semiconductor (TC bonders + hybrid bonders), asmt (TCB for HBM stacking)
- Key context: SK Hynix's HBM4 mass production timing in Feb 2026 is the single most important supply-chain date for the 2026-2027 AI GPU cycle.