# How Bottleneck Hunting Identifies Supply Chain Opportunities in AI Berkshire

> Discover how bottleneck hunting in AI Berkshire uncovers supply chain opportunities. This systematic pipeline scores shortages and maps them to undervalued public companies for investable ideas.

- Repository: [Xbt Lin/ai-berkshire](https://github.com/xbtlin/ai-berkshire)
- Tags: deep-dive
- Published: 2026-06-30

---

**Bottleneck hunting is a systematic six-step pipeline that converts supply-chain analysis into investable ideas by scoring component shortages and mapping them to undervalued public companies.**

The **AI Berkshire** repository implements bottleneck hunting as a markdown-based skill that analyzes capital-intensive super-trends and decomposes them into physical supply-chain layers. This methodology, defined in [`skills/bottleneck-hunter.md`](https://github.com/xbtlin/ai-berkshire/blob/main/skills/bottleneck-hunter.md), enables investors to identify structural supply constraints before they are fully priced by the market.

## The Six-Step Bottleneck Hunting Pipeline

The workflow operates as a narrowing funnel from macro trends to specific equities. Each step is encoded in the skill file and executed by the AI Berkshire command router.

### 1. Super-Trend Confirmation

The pipeline begins by validating long-run, capital-intensive trends against four strict filters. A trend must demonstrate **duration ≥ 3 years**, **physical hardware requirements**, **global CAPEX > $50 billion/year**, and **demand growth exceeding supply growth**. The approved trend list is stored in the skill file and refreshed on every invocation.

### 2. Supply-Chain Physical Decomposition

Each confirmed trend is broken into a **layered-tree model** (Layer 0 → Layer 4) that separates end-products from core components, sub-components, upstream materials, and infrastructure. For example, the AI infrastructure trend decomposes from GPU → optical modules → InP substrates, with each layer hard-coded in the skill's decomposition tables.

### 3. Bottleneck Identification

Every Layer 2-3 element is scored against six hard criteria: **concentration**, **expansion lead-time**, **substitutability**, **utilization**, **demand-growth**, and **validation cycle**. A **6-point rubric** generates S/A/B ratings where S indicates single-point-of-failure risk, A indicates severe constraints, and B indicates manageable supply.

### 4. Company Screening

High-rated bottlenecks are mapped to publicly listed suppliers using web searches for "{bottleneck} supplier listed company". The system applies fast-filter rules requiring companies to be publicly listed, derive **≥ 30% revenue from the bottleneck**, maintain **market cap < $10 billion**, and demonstrate **daily turnover > $1 million**.

### 5. Cross-Verification

Each candidate undergoes validation through a **positive/negative verification matrix**. The skill checks for client contracts, CAPEX guidance, price hikes, and capacity constraints while simultaneously running a "芒格式否定" (Munger-style negation) checklist to surface bearish arguments and disconfirming evidence.

### 6. Output Generation

The final step emits a concise **opportunity card** containing valuation checks (PS/PE thresholds, TAM-ratio, 10-year return tests) and a signal-strength rating (★★★★★ → ★). The output format is fixed in the skill file's "报告模板" (report template) section.

## S/A/B Bottleneck Scoring Methodology

The rating system quantifies supply-chain fragility using objective metrics.

- **S-Rated (Single-Point-of-Failure)**: Concentration ≤ 2 suppliers, expansion lead-time > 18 months, zero substitutability, utilization > 90%, demand growth 2× supply growth, validation cycle > 12 months.
- **A-Rated (Severe)**: Concentration 3-4 suppliers, expansion lead-time 12-18 months, limited substitutability, utilization 80-90%.
- **B-Rated (Manageable)**: Concentration > 4 suppliers, expansion lead-time < 12 months, available substitutes, utilization < 80%.

This rubric is defined in the markdown table within [`skills/bottleneck-hunter.md`](https://github.com/xbtlin/ai-berkshire/blob/main/skills/bottleneck-hunter.md) and parsed dynamically during execution.

## Practical Usage Examples

### CLI Invocation via Claude Code

Researchers trigger the workflow using slash commands:

```markdown
/bottleneck-hunter AI基础设施

```

This command initiates the six-step pipeline for the "AI infrastructure" super-trend and returns a formatted markdown report with investment candidates.

### Sample Output Structure

The skill generates standardized opportunity cards:

```markdown

# 瓶颈猎手 — 2026-06-30 14:00

## 明确标的

### 雪峰科技（688123）— 光模块供应链瓶颈

**瓶颈定位**: Layer 2，光模块，S级瓶颈  
**财务快照**: 市值 $8.2 bn / 年收入 $1.1 bn / PS 7.5x / PE 22x  
**估值检查**: 绿灯（PS < 10 x，PE < 30 x）  

**看多逻辑**  
1. 供给集中度≤2家 → 议价能力强  
2. 客户签约已公布 → 收入可验证  

**看空逻辑**  
1. 替代技术（硅光子）研发进度快  

```

### Programmatic Integration

For automated research pipelines, call the skill via the Codex prompt endpoint:

```python
import requests

payload = {
    "skill": "bottleneck-hunter",
    "arguments": "AI基础设施"
}
r = requests.post("http://localhost:8000/run-skill", json=payload)
print(r.json()["report"])

```

The endpoint reads the same markdown template and returns the rendered report as a JSON field, enabling integration with quantitative trading systems.

## Core Implementation Files

The bottleneck hunting capability is distributed across four key locations in the **xbtlin/ai-berkshire** repository:

- **[`skills/bottleneck-hunter.md`](https://github.com/xbtlin/ai-berkshire/blob/main/skills/bottleneck-hunter.md)**: Contains the complete workflow definition, scoring rubrics, layer decomposition trees, and report templates.
- **[`codex-prompts/bottleneck-hunter.md`](https://github.com/xbtlin/ai-berkshire/blob/main/codex-prompts/bottleneck-hunter.md)**: Maps the markdown skill to CLI-compatible commands for execution.
- **`reports/bottleneck-map/`**: Storage directory for daily and master bottleneck maps generated by the pipeline.
- **[`README.md`](https://github.com/xbtlin/ai-berkshire/blob/main/README.md)**: Provides high-level project overview and skill installation instructions.

## Summary

- **Bottleneck hunting** in AI Berkshire follows a rigid six-step pipeline from trend confirmation to opportunity generation.
- The methodology uses a **6-point rubric** to classify supply constraints as S, A, or B ratings based on concentration and lead-time metrics.
- Company screening applies **financial filters** (market cap < $10bn, revenue concentration ≥ 30%) to identify asymmetric opportunities.
- All logic is implemented in **[`skills/bottleneck-hunter.md`](https://github.com/xbtlin/ai-berkshire/blob/main/skills/bottleneck-hunter.md)** and executed via slash commands or API calls.
- Output includes **structured opportunity cards** with built-in valuation checks and bearish argument matrices.

## Frequently Asked Questions

### What distinguishes an S-rated bottleneck from an A-rated one?

An **S-rated bottleneck** represents a single-point-of-failure with supplier concentration ≤ 2, expansion lead-times exceeding 18 months, and utilization above 90%. An **A-rated bottleneck** indicates severe but manageable constraints with 3-4 suppliers and 12-18 month lead-times. The distinction determines position sizing and urgency in the final investment recommendation.

### How does the system verify bottleneck investment candidates?

The skill implements a **positive/negative verification matrix** requiring at least two independent sources for validation, such as client announcements, CAPEX guidance, or industry reports. Simultaneously, it runs a "芒格式否定" checklist to actively seek disconfirming evidence and alternative technologies that could obsolesce the bottleneck component.

### What financial filters are applied during company screening?

The screening layer filters for **publicly listed companies** with **≥ 30% revenue exposure** to the specific bottleneck, **market capitalization below $10 billion**, and **daily turnover exceeding $1 million**. These constraints target liquid, small-cap pure-plays with sufficient asymmetry potential.

### Where is the bottleneck hunting logic defined?

All workflow logic, scoring rubrics, and report templates are defined in **[`skills/bottleneck-hunter.md`](https://github.com/xbtlin/ai-berkshire/blob/main/skills/bottleneck-hunter.md)** within the AI Berkshire repository. This markdown file is parsed by the command router and executed whenever a user invokes the `/bottleneck-hunter` skill.