Skip to content

P4-C2 · Samsung Loses HBM Market Share (5-Step Bottleneck Real Demo)

Core Takeaway

3 companies in the same industry (SK Hynix / Micron / Samsung) saw ±150% stock divergence in 1 year — micro differences = massive divergence.

Real Public Case Study — Using Part 3 Tools to Analyze 5 Real Events

P4-C2. All based on public data (company earnings reports + earnings transcripts + public industry news).

1. Event: Full Year 2024 HBM Stock Divergence of ±150% Across 3 Companies

Company 2024 Stock YoY Key Event
SK Hynix +130% HBM3e first to qualify for NVDA H200, supplying 70%+ share
Micron +60% HBM3e volume production 2024/02 (industry first, 24GB 8-high in H200), full-year ramp + gain share
Samsung -20% Slow HBM qualification + worker strike + weak smartphone market

All 3 make the same product (HBM high-bandwidth memory), yet stock prices diverged by 150%. Why?

2. Using P3-C3's 5-Step Bottleneck Process to Verify Each Step

2.1 Step 1 — Identify (Q1 2023)

Read NVDA earnings call: Jensen mentions "supply constrained on HBM". → Bottleneck signal.

Read SK Hynix earnings: HBM backlog doubles. → Demand confirmation.

→ If you noticed the keyword "supply constrained on HBM" in Q1 2023, you were 6 months ahead of everyone.

2.2 Step 2 — Track Supply

Public data (3 companies' earnings calls):

Company HBM Capacity Expansion Plan NVDA Qualification Status
SK Hynix HBM3e ramp in Q1 2024 Already qualified, primary supplier
Micron 2024/02 volume production (24GB 8-high in H200) Qualification passed
Samsung Target Q3 2024 Failed qualification multiple times

→ Key insight: Samsung's slow qualification = missing this NVDA wave.

2.3 Step 3 — Track Demand

Hyperscaler capex (full year 2024):

→ Big 4 hyperscaler 2025 actual capex ~$270-290B (annualized; AI-only scope vs total scope must be distinguished). GPU demand → HBM demand.

Simple math (caveat: capex CANNOT be directly divided by GPU ASP — hyperscaler capex includes land / buildings / power / cooling / networking / CPUs / storage / ASICs / leases / construction): - Estimate AI accelerator / server bucket ~40-50% of ~$290B = ~$120-145B AI server - NVDA accounts for ~80% of AI accelerators = ~$95-115B NVDA share - At ~$40K/GPU ASP (system-level, not bare-GPU) → ~2.4-2.9M GPU/year (vs the old "3.2M" using bare-chip ASP; numbers depend on ASP / mix assumptions — the point is strip out non-GPU buckets BEFORE dividing by ASP) - HBM demand (split by generation): - H100/H200 sensitivity: ~2.4-2.9M × 80-141GB HBM = ~200-400M GB/year (NVDA H100 80GB / H200 141GB HBM3e per official spec) - Blackwell sensitivity: ~2.4-2.9M × ~180-192GB HBM = ~430-560M GB/year (DGX B200 platform ~180GB/GPU; B100/B200 module configs differ) - (R10 Fix 5: Blackwell 192GB is a next-gen separate range, not within the 80-141GB H100/H200 range)

vs 2024 global HBM capacity ~1B GB → Demand / Supply = 2.4x. Extreme shortage.

2.4 Step 4 — Track Price + Lead Time

Public signals:

  • HBM3e prices rose 25-40% in 2024 (various industry news)
  • NVDA H100 lead time 36-52 weeks (CRWV public interview)
  • SK Hynix Q2 2024 earnings: HBM fully sold out through 2026

→ Extreme shortage, sustained for multiple quarters.

2.5 Step 5 — Sentiment

  • Sell-side upgrades: SK Hynix multiple target price increases
  • 13F / holding trackers Q1 2024: Tiger / Coatue added SK Hynix-linked securities (e.g. ADR/GDR/related US-listed instruments) — note: SK Hynix KRX 000660 common shares are NOT 13(f)-reportable (Section 13(f) only covers US-listed securities); specific instrument/CUSIP is required (R10 Fix 6)
  • Stock price: SK Hynix Q1 2024 +60% YTD

→ By now, retail investors are just seeing the news.

3. Why Samsung Lost Market Share — A Triple Blow

3.1 Technical Reason: Failed Qualification

NVDA HBM qualification requirements: - Bandwidth ≥ HBM3e spec - High yield (NVDA cannot tolerate batch variation) - Long-term reliability (24/7 100% load)

Samsung submitted multiple times in 2023-2024 and failed each time. SK Hynix passed 1 year earlier, Micron followed.

3.2 Strike

In July 2024, Samsung workers went on strike for 25 days — Samsung stated no production disruption; market worried about HBM capacity signal (2026/05 planned 18-day strike suspended; market estimates of 18-day strike: ~3-4% DRAM / 2-3% NAND global supply). 8-12% impact figure lacks reliable source.

3.3 Weak Smartphone Market

Samsung memory isn't just HBM — DRAM / NAND is largely for phones / servers. The weak 2024 smartphone market dragged down the overall memory cycle, diluting HBM gains.

→ Triple blow stacked, all visible in public data.

4. Lesson — Micro Differences Within the Same Industry Determine 100% of Stock Performance

Lesson How to Use in Your Thesis
Qualification is a binary signal 1 company passes vs fails = ±100% stock difference
Focus on micro comparisons within an industry Don't buy sector ETFs, cherry-pick individual stocks
Monitor public earnings call keywords "qualify" / "supply constrained" / "ramp" are gold
Strikes / factory outages are leading signals News appears 1-3 months before earnings

Key point: Retail investors see "memory chips up" and buy the ETF — but the ETF includes Samsung. Result: SK Hynix +130% vs Samsung -20%, average returns are mediocre. Picking the right single stock is the source of returns.

5. What Your Thesis Should Learn

  • Don't buy "memory" sector ETFs — buy SK Hynix or Micron, not Samsung
  • Monitor qualification status — keywords in NVDA earnings + 3 supplier earnings
  • 2026 HBM4 cycle repeats: SK Hynix has already secured the ramp, Samsung is still qualifying. Next 6-12 months, watch which company passes first

6. Next → P4-C3

You've seen the supply chain wildcard. Next is the government capex wildcard — Stargate $500B.

→ P4-C3 · Stargate $500B Beneficiary Chain.

7. Further Reading (this chapter — Samsung missing H200 HBM3E qualification)

All free sources, primary earnings transcripts / IR / industry research preferred

Primary earnings / IR (company original disclosures):

Primary announcements / product specs:

TrendForce / industry research (free reports):

Reuters / primary news (free):

Wikipedia (background):

Podcasts (1-2 hr):

  • Acquired · TSMC (2021) — doesn't cover HBM directly but explains foundry / packaging / qualification culture in semiconductors, useful for understanding why qualification is so hard

Pair with chapter self-check: After reading any 3-4 of the above, you should be able to answer the self-check more confidently — especially on "why SK Hynix ramped first" and the real-data argument for "Samsung's triple blow (qualification / yield / customer)".