5 Introductory Cases (Building Intuition)¶
Each case takes about 5 minutes to read — no need to understand jargon. First, build intuition for "what happens in the AI industry." A detailed deep-dive version is in Layer 4.
Case 1: NVIDIA Loses $593B Overnight¶
January 2025 — China's DeepSeek releases the V3 model, with performance close to OpenAI's GPT-4, but reportedly using only 1/10 of the compute for training.
The next day at US market open — NVIDIA drops 17% in a single day, losing $593B in market cap (the largest single-day loss in history).
Why: The market feared, "If AI training efficiency improves 10x, are future GPU demands overestimated?"
Aftermath: Within a week, NVDA recovered 60-80%. The market later realized: DeepSeek's breakthrough wasn't "we don't need as many GPUs," but rather "compute is used more efficiently → the application layer will explode → total compute demand will actually be larger" (Jevons Paradox).
Lesson: A single news event → short-term price volatility → but if the thesis hasn't truly changed, the long-term reverts to fundamentals.
Case 2: Samsung vs SK Hynix — The HBM Battle¶
Full Year 2024 — AI GPUs (NVIDIA H100/H200) require a special high-bandwidth memory called HBM (High Bandwidth Memory). Only three companies globally can produce it: SK Hynix / Micron / Samsung.
SK Hynix qualified early and became NVIDIA's main supplier. Micron followed. Samsung kept failing to qualify, losing market share.
2025 — Samsung workers go on strike (at a memory factory), further impacting HBM production capacity.
Hynix stock price: +130% YoY
Samsung stock price: -20%
Lesson: Within the same industry, micro-level differences like "whether technical specs pass NVIDIA's qualification" can determine a 100%+ gap in stock price performance.
Case 3: Stargate Phase 1 — $500B AI Infrastructure¶
January 2025 — After Trump's inauguration, OpenAI / Oracle / Microsoft / SoftBank jointly announce the Stargate project: a $500B investment over 5 years to build AI data centers (first stop in Abilene, Texas).
Beneficiary tickers (7 days after the announcement): - ORCL +15% (Oracle Cloud running OpenAI workloads) - CRWV (CoreWeave) +20% - VRT (Vertiv, data center electrical) +12% - CEG (Constellation, nuclear PPA) +10%
Lesson: AI isn't just a chip company story — electrical contractors (ETN/HUBB) / data center REITs (EQIX/DLR) / nuclear power companies (CEG/VST) / gas turbines (GEV) are all "selling shovels."
Case 4: Export Controls Revive Huawei / SMIC¶
2022-2023 — The US imposes export controls on AI chips to China (H100/H200 cannot be sold to China). The market thought this was bearish for NVIDIA (losing a market).
Actual development: - NVIDIA loses ~25% of revenue from China - But other overseas customers fill the gap → total revenue still rises - Meanwhile, **Huawei (Ascend 910B) + SMIC** begin to rise as "China's version of AI chips," with SMIC revenue jumping 30%+
Lesson: Regulatory actions have a 6-24 month lag. In the short term, the "obviously harmed party" may be offset by other factors, and local substitutes in the restricted region become a new thesis.
Case 5: 13F Position Changes ≠ Short-Term Price Signal¶
May 2026 — 13F reports are disclosed (institutional Q1 holdings): - Bridgewater reduces CIEN by -32% - Lone Pine reduces CIEN by -60% - Citadel increases CIEN by +39%
Our thesis at the time: bull (Citadel/Point72 increasing positions is a positive signal)
7 days later: CIEN -4.6% (net institutional selling has a bigger impact than net buying)
14 days later: CIEN +3% (reverts to fundamentals, earnings expectations improving)
Lesson: 13F is a quarter-end snapshot, not current holdings. It shows what happened 3 months ago and has no direct causal relationship with next week's stock price. Its value lies in observing institutional thesis trends, not as a timing signal.
Common Lessons from These 5 Cases¶
| Case | Core Lesson |
|---|---|
| DeepSeek sell-off | Short-term price reaction ≠ long-term fundamental change |
| Samsung HBM | Micro-level differences within the same industry = huge stock price divergence |
| Stargate | "Shovel sellers" have better odds than "gold miners" |
| Export controls | Regulatory impact has a lag; there's always a beneficiary |
| 13F | Data ≠ signal; look at the timestamp and causality |
Next step → AI Industry Chain Overview