Glossary

Key terms for understanding AI markets and investor holdings — concise, bilingual, and easy to reference.

Investing principles

Moat
A durable advantage that fends off competition and sustains high returns — network effects, switching costs, scale, proprietary data, or ecosystem lock-in.
Circle of competence
The domains you genuinely understand well enough to judge their business and competition. Investing only within it is a core Buffett discipline.
Margin of safety
The cushion between price paid and intrinsic value. The wider it is, the more room for error and bad luck.
Drawdown
The peak-to-trough decline. AI is volatile with deep drawdowns; position sizing decides whether you can hold through them.
Dollar-cost averagingDCA
Investing a fixed amount on a fixed schedule to smooth cost and reduce timing risk — suited to long-term themes.
Core-satellite
A portfolio approach: a stable core in broad index or platform leaders, plus small satellites for high-beta themes.

Markets & data

13F filing13F
A quarterly SEC filing where large U.S. institutions disclose their U.S. equity holdings. Useful for tracking the greats, but lagged (~45 days after quarter-end).
Capital expenditureCapex
Spending on long-term assets (data centers, chips). AI’s massive capex is the core driver of compute and infrastructure demand.
Price-to-earningsP/E
Share price divided by earnings per share — the price paid per unit of profit. High multiples are very sensitive to any slowdown.
Free cash flowFCF
Operating cash flow minus capex — the cash a business can truly deploy, a gauge of earnings quality.

AI & compute

Picks and shovels
Selling the tools rather than the end product — in AI, the chips, compute, and infrastructure with the most certain cash flows.
Custom silicon
In-house AI accelerators from hyperscalers (Trainium, TPU, Maia) that erode merchant GPU share over time.
High-bandwidth memoryHBM
High-speed memory paired with AI accelerators; tight supply lifts volume and price, benefiting names like Micron.
Training / Inference
Training teaches a model from data (compute-intensive); inference serves the trained model to users (at scale, with ongoing compute).
Large language modelLLM
Large AI models trained on vast text (GPT, Gemini, Ernie) — the foundation of this AI application wave.