Samsung Electronics (HBM Division)

The #2 HBM supplier with a comeback story — Samsung cleared NVIDIA HBM3e qualification in late 2025 and is racing SK Hynix for HBM4 market share. The only credible alternative supplier for AI GPU HBM, but yield challenges have kept it behind SK Hynix.

1. Core Product / Service

Samsung Electronics' memory division competes in HBM alongside sk-hynix and Micron. Key differentiator: Samsung is the only HBM maker that also has its own foundry + OSAT capabilities, enabling a fully internal chip supply chain — theoretically.

For AI datacenter HBM specifically:

• HBM3e — Samsung's 8-layer and 12-layer HBM3e products. Samsung reportedly failed initial NVIDIA qualification tests in 2024 (12-Hi variant), but later cleared qualification for 8-layer HBM3e in 2025 [8]. The 12-Hi HBM3e qualification failure gave sk-hynix a near-monopoly on NVIDIA HBM supply through 2025.
• HBM4 — Samsung adopted a more aggressive design path: 1c-class DRAM (sub-10nm) and 4-nm logic base die, targeting >11 Gb/s per pin — potentially higher performance than SK Hynix's HBM4. Mass production was delayed to 2026 due to yield challenges (Aug 2025) [source: SemiWiki/DigiTimes]. Samsung and SK Hynix both began HBM4 mass production in February 2026 [Silicon Analysts].
• Google TPU relationship — Samsung supplies >60% of Google's HBM3e for TPU systems via broadcom (2025 data) [1]. This is a separate supply relationship from NVIDIA.

2. Target Users & Pain Points

• NVIDIA — after clearing HBM3e qualification, Samsung is in contention for HBM4 orders alongside SK Hynix
• Google — primary HBM supplier for TPU via broadcom custom ASIC [1]
• Broadcom — Samsung's Broadcom relationship is stronger than its NVIDIA relationship; Broadcom uses Samsung for Google TPU HBM and some other ASIC programs
• Potential future: if Samsung's HBM4 yield improves, it can displace some sk-hynix share at NVIDIA

Pain solved: NVIDIA wants a second HBM supplier besides SK Hynix — purely from a supply security standpoint. Having 60-70% of HBM supply with one vendor is a strategic risk NVIDIA would prefer to reduce. Samsung's HBM4 qualification is therefore a top priority for NVIDIA.

3. Competitive Landscape

Samsung's competitive position: the only credible alternative to SK Hynix for AI GPU HBM. If Samsung gets HBM4 right, it can win back significant NVIDIA share. If yield issues persist, it remains behind. The gap is not technology — Samsung's HBM4 specs are competitive — but manufacturing execution and yield.

4. Unique Observations

• Samsung cleared NVIDIA HBM3e qualification in late 2025 for 8-layer HBM3e, ending the period where Samsung had failed NVIDIA qualification tests [8]. The 12-Hi HBM3e struggle cost Samsung significant market share in 2024-2025.
• HBM4 strategy — higher-performance path: Samsung uses 1c DRAM (more advanced than SK Hynix's 1b nm) and 4-nm base die (vs TSMC 12nm for SK Hynix). If Samsung can achieve good yield at this specs, HBM4 could have performance advantages.
• Samsung HBM4 production delayed to 2026 due to yield challenges, while SK Hynix hit mass production Feb 2026 [SemiWiki, Aug 2025]. Both started HBM4 mass production in February 2026 [Silicon Analysts], suggesting Samsung closed the gap.
• Reuters: Samsung in talks with NVIDIA to supply HBM4 (Oct 2025) [7] — this confirms NVIDIA is actively seeking Samsung as a second HBM4 source.
• The Broadcom-Google relationship is Samsung's quiet strength: Samsung supplies >60% of Google TPU HBM via Broadcom [1]. This gives Samsung a large, stable AI HBM revenue stream outside the NVIDIA supply chain.
• Yield risk is the persistent issue: Samsung's foundry business has had well-documented yield issues across multiple advanced nodes (3nm, 4nm). The same pattern appears in HBM — 12-Hi HBM3e yield was insufficient for NVIDIA qualification. Whether HBM4 yield is production-ready is the central question for Samsung's AI story.
• Samsung-TSMC CoWoS interaction: Samsung does its own foundry + packaging; NVIDIA uses TSMC CoWoS. So Samsung doesn't directly compete with TSMC for CoWoS. But Samsung's HBM feeds into TSMC CoWoS lines (or Samsung's own packaging lines). The relationship is complex.

5. Financials / Funding

• Listed: KRX: 005930; Samsung Electronics total; HBM division not separately disclosed
• Note: Samsung Electronics is a conglomerate; HBM financials are embedded in the Device Solutions division. Market cap: Samsung Electronics ~$350B+.
• HBM3e qualification impact: Clear revenue positive in 2026 as Samsung begins shipping HBM3e to NVIDIA.
• HBM4 revenue: potential to begin in 2026 if yield allows.

6. People & Relationships

• NVIDIA: qualification status (8-layer HBM3e cleared, HBM4 in negotiation)
• Broadcom: Google TPU HBM supplier; stronger Broadcom relationship than NVIDIA relationship
• TSMC: complex — Samsung foundry competes with TSMC; Samsung HBM is sold to TSMC packaging lines or Samsung's own packaging
• Key strategic context: Samsung is the only company that has foundry + memory + packaging all under one roof. If all three execute, it's a vertically integrated AI chip competitor. Currently, execution at the memory HBM and foundry nodes has been uneven.