Chapter 06 / 16
Scoring and Classification
瓶颈重要性、垄断潜力、共识定价和发现差必须分别评价。
Scoring Model
Score each company on two separate dimensions:
Hard Gate: Monopoly / Oligopoly Potential
Do not select a stock for the Top 10 unless it either already has, or has a credible path toward, monopoly-like or oligopoly-like control over a bottleneck layer.
This must be derived by reasoning from market structure, not copied from search results.
Required reasoning chain:
- Define the narrow bottleneck market.
Example: not "AI chips", but "DDR5 RCD/MRCD/MDB for server memory modules" or "PCIe/CXL retimers for AI servers".
- Identify why customers cannot easily avoid this layer.
Examples: standards requirement, qualification cycle, platform compatibility, reliability risk, switching cost, supply scarcity.
- Identify why competitors cannot easily enter.
Examples: SerDes expertise, JEDEC/CXL/PCIe standards participation, hyperscaler design-in, long validation cycles, IP portfolio, customer certification.
- Identify the company's position in that bottleneck.
Examples: current share, design wins, platform inclusion, customer concentration in leading buyers, ecosystem role.
- Identify the failure mode.
Examples: standard changes, customer self-development, second-source pressure, price compression, replacement by another protocol.
Monopoly potential score:
- 1: Commodity supplier; no structural control.
- 2: Niche supplier but easy to substitute.
- 3: Oligopoly candidate; some switching cost or certification moat.
- 4: Strong oligopoly position in a narrow bottleneck market.
- 5: De facto monopoly or unavoidable platform layer with durable pricing power.
Top 10 rule:
monopoly_potential_score >= 3.5
If a company has a strong AI narrative but a weak monopoly path, label it avoid_narrative_only or cycle_beta, not a core pick.
A. Bottleneck Score
Score 1-5 on:
- Bottleneck severity: Is this needed for AI scaleout?
- Supply rigidity: Is supply hard or slow to expand?
- Customer urgency: Are hyperscalers, AI clouds, or enterprises forced to buy?
- Evidence velocity: Are new deals, filings, capex, or hiring accelerating?
Formula:
bottleneck_score = bottleneck_severity supply_rigidity customer_urgency * evidence_velocity
B. Consensus / Priced-In Score
Score 1-5:
- 1: Little market attention; mostly absent from mainstream AI narratives.
- 2: Some specialist attention; limited sell-side / media coverage.
- 3: Recognized by sector investors, but still debated or inconsistently priced.
- 4: Broadly recognized; multiple broker reports, strong narrative, valuation already reflecting growth.
- 5: Fully consensus AI trade; crowded, expensive, and already treated as obvious beneficiary.
Inputs for this score:
- Stock move versus sector over 3 / 6 / 12 months.
- Valuation expansion versus its own history.
- Frequency in broker notes, financial media, social media, and fund letters.
- Whether AI exposure is already explicit in earnings calls.
- Whether revenue acceleration is already visible or still only an option.
C. Discovery Gap
The opportunity target is not simply the highest bottleneck score. It is:
discovery_gap = bottleneck_score_normalized - consensus_priced_score
Interpretation:
- High bottleneck + high consensus: probably real, but already expensive.
- High bottleneck + low consensus: best deep-dive candidate.
- Low bottleneck + high consensus: avoid unless short-term trading only.
- Low bottleneck + low consensus: ignore unless new evidence appears.
Current Qualitative Consensus Map
As of 2026-06-20, based on recent market and industry signals:
| Theme | Bottleneck | Consensus / priced-in | Discovery gap | Notes |
|---|---|---|---|---|
| Nvidia GPU / full-stack AI compute | 5 | 5 | 0 | Real bottleneck, but fully consensus. Use as demand signal, not hidden opportunity. |
| HBM leaders: SK hynix, Micron, Samsung | 5 | 4-5 | 0-1 | Very real, but now widely recognized. Still cyclical upside possible, less undiscovered. |
| Nearline HDD: Seagate, Western Digital | 4 | 3-4 | 1 | AI storage demand is being discovered; not hidden anymore, but less crowded than HBM. |
| CXL / memory pooling: Astera, Rambus, 澜起科技 | 4 | 2-3 | 1-2 | Better non-consensus layer: key if memory shifts from fixed server DRAM to pooled memory. |
| DDR5 interface / retimer: 澜起科技, Rambus, Astera | 4 | 2-3 | 1-2 | More attractive than generic memory modules because it sits in the control/interconnect layer. |
| Enterprise SSD / NAND controllers: Silicon Motion, SanDisk, 江波龙, 佰维存储 | 3-4 | 3 | 0-1 | Good cycle trade, but more exposed to commodity pricing and inventory swings. |
| AI host CPU attach: AMD, Arm, Nvidia Grace/Vera | 4 | 3-4 | 0-1 | Becoming visible; still worth tracking because GPU rack architectures force CPU attach. |
| Intel CPU / foundry / packaging turnaround | 3-4 | 3 | 0-1 | Potentially large, but execution risk is high; separate turnaround thesis from AI bottleneck thesis. |
| A-share国产CPU: 海光信息, 龙芯中科 | 3-4 | 3-4 | 0-1 | More policy/import-substitution than pure AI bottleneck; valuation often already anticipates this. |
| AI server OEM/ODM: 浪潮信息, 中科曙光, 工业富联 | 3 | 3-4 | -1-0 | Demand is obvious, margins may be competed away; watch order quality and profit capture. |
Automation Label
Each company should receive one of these labels:
crowded_winner: real winner, but already heavily priced.watch_for_pullback: good business, weak margin of safety at current consensus.underdiscovered_bottleneck: best target for deep research.cycle_beta: mostly memory / hardware cycle exposure.execution_option: upside depends on company execution, not only industry bottleneck.avoid_narrative_only: AI story exists, but bottleneck capture is unclear.
Initial labels:
| Company | Market | Label | Rationale |
|---|---|---|---|
| Nvidia | US | crowded_winner | Strongest AI infrastructure position, but fully consensus. |
| Micron | US | watch_for_pullback | HBM/DRAM/NAND bottleneck is real, but memory boom is now widely recognized. |
| Seagate | US | watch_for_pullback | Nearline HDD scarcity is increasingly recognized; still less crowded than HBM. |
| Western Digital | US | watch_for_pullback | Similar HDD/storage thesis; monitor pricing discipline and supply commitments. |
| Astera Labs | US | underdiscovered_bottleneck | CXL/retimer/connectivity layer is strategically important, but valuation may already embed some scarcity premium. |
| Rambus | US | underdiscovered_bottleneck | Memory-interface IP and chips may benefit from DDR5/CXL without taking full commodity memory risk. |
| AMD | US | watch_for_pullback | CPU attach and AI accelerator upside are real, but increasingly consensus. |
| Arm Holdings | US | watch_for_pullback | Server Arm attach is real, but valuation often prices a broad architecture win. |
| Intel | US | execution_option | Foundry/packaging/CPU recovery has upside, but thesis depends on execution. |
| 澜起科技 | A-share | underdiscovered_bottleneck | DDR5 interface, PCIe retimer, CXL MXC; better positioned in memory interconnect than generic storage names. |
| 海光信息 | A-share | watch_for_pullback | Domestic CPU leader; strong policy/国产替代 narrative already partly priced. |
| 中科曙光 | A-share | cycle_beta | AI server/system demand beneficiary; profit capture needs verification. |
| 浪潮信息 | A-share | cycle_beta | High AI server relevance, but more competitive hardware margin profile. |
| 江波龙 | A-share | cycle_beta | Storage cycle and module upside, but more exposed to commodity swings. |
| 佰维存储 | A-share | cycle_beta | Similar storage cycle beta; needs evidence of durable enterprise/AI mix. |
Primary automation sort:
research_priority = monopoly_potential_score + discovery_gap + evidence_velocity - valuation_heat
Automation should surface:
- New high-priority company mentions.
- Companies with rising evidence velocity.
- New bottleneck terms that appear across multiple sources.
- Disagreements between investor narrative and operational evidence.
- Companies where consensus score rises faster than evidence velocity, which may mean the trade is becoming crowded.