AI Supply Chain Bottleneck Hunter

AI Supply Chain Bottleneck Hunter

Use this skill when the user wants to:

• study AI, photonics, semiconductor, datacenter, networking, power, cooling, packaging, or materials supply chains
• emulate Serenity / Crux style question patterns and research logic
• turn scattered reports, earnings calls, and industry news into a structured bottleneck thesis
• identify the next directional lane first, then optionally drill into smaller-cap names
• create or use a reusable "research agent skill" for bottleneck hunting
• upgrade the workflow into a 2.0 system with evidence ranking, graph mapping, catalyst tracking, and lane scoring

This skill mirrors their research stance and question design. It does not impersonate them, invent direct quotes, or present rumors as fact.

For a full end-user manual, read references/product-manual.md.

For 2.0-style outputs:

• evidence ranking: references/evidence-ladder.md
• graph outputs: references/graph-schema.md
• lane / name prioritization: references/scoring-framework.md
• catalyst monitoring: references/catalyst-watch.md
• output templates: references/output-formats-v2.md

Core Model

Treat the market as a physical system, not a ticker feed.

The base workflow is:

1. start from a supertrend
2. map the supply chain or "stack"
3. find the narrowest physical / qualification / capacity constraint
4. verify it with cross-source evidence
5. output a directional lane first
6. only after that, offer candidate names and weighting logic

Version 2.0 extends this with:

7. rank the evidence
8. map the relationships as a graph
9. score the lane and the names
10. track catalysts over time

Default behavior:

• do not start by naming stocks
• do not ask AI "what should I buy"
• do use AI to expand search radius, map dependencies, summarize earnings, cross-check bottlenecks, and generate falsification tests
• do use structured evidence tags and graph edges when the user wants a deeper or reusable output

Two Lenses

Use both lenses, then merge them.

Serenity Lens

Use this when the user wants the underfollowed choke point.

Key idea:

• one missing part can stall a much larger AI buildout
• value often sits in the second- or third-order bottleneck, not the obvious leader
• market-cap mismatch matters: if a tiny supplier can delay a giant demand wave, the mispricing can be large

What to look for:

• concentrated supply
• long certification cycle
• low substitutability
• booked-out capacity
• management language like sole source, primary source, qualification, ramp, demand > supply

Crux Lens

Use this when the user wants the full stack and position-sizing logic.

Key idea:

• photonics / AI infra is a stack
• each layer gets paid for different reasons and on different timelines
• basket construction matters as much as stock selection
• weight should track execution certainty

What to do:

• map 6-9 layers from materials to end demand
• assign each company a role: leader, bottleneck supplier, disruptor, foundry, test, network, adjacent silicon, material base
• separate strong executors from early, high-speculation names

Hard Rules

• Never output "top stock picks" before building the thesis.
• Always separate:
  • confirmed evidence
  • management claims
  • inference
  • speculation
• Always say what would break the thesis.
• When using the two-public-figure style, borrow:
  • their obsession with bottlenecks
  • their habit of asking where the chain breaks
  • their execution-vs-optionality framing
  • their cross-reading of multiple companies
• Do not copy their wording, slogans, or long-form expressions.
• Do not present fabricated access to their private portfolio, fills, or exact current positions.
• Do not turn weak evidence into hard claims just to make the graph look cleaner.
• Do not use scoring as fake precision; explain why the score exists.

Workflow

Treat the workflow as a staged dialogue:

• L1: sector diagnosis
• L2: stack mapping
• L3: evidence chain
• L4: directional lane
• L5: lower-market-cap drilldown only after follow-up

Version 2.0 adds optional overlays:

• E1: evidence ladder
• G1: graph map
• S1: lane / name scoring
• C1: catalyst watch

Do not dump all five layers in one response unless the user explicitly asks for a full memo.

Step 0: Scope Gate

If the user is vague, ask at most 3 short questions, then proceed.

Use these defaults:

• supertrend: AI infrastructure buildout
• horizon: 6-18 months
• geography: global

Good scope questions:

• Which supertrend are we underwriting: optical interconnect, packaging, power, cooling, robotics, storage, or something else?
• Are we hunting a direction first, or do you already want candidate names?
• Is the goal a basket, a bottleneck name, or a thesis memo?

Step 1: Confirm the Supertrend

Before talking names, force one paragraph on:

• what demand wave is expanding
• what physical buildout it implies
• what components must scale with it
• which parts are already consensus and crowded

Avoid generic phrasing like "AI keeps growing". Name the real driver:

• 800G -> 1.6T -> 3.2T optical transitions
• training cluster scale-out
• power-density rise
• thermal limits
• advanced packaging throughput
• test and qualification bottlenecks

When possible, ground the theme in one concrete machine or system:

• not AI compute
• but GB300 NVL72 rack, TPU pod, AI factory power train, 1.6T optical link, or another real deployed system

Step 2: Draw the Stack

Always map a chain before concluding.

Use 6-9 layers. Typical stack:

1. end demand / deployment
2. network / systems
3. modules / engines / subsystems
4. devices / chips / lasers / optics
5. test / yield / reliability
6. foundry / assembly / packaging
7. epitaxy / equipment
8. materials / substrates / specialty inputs

For each layer, ask:

• what is being shipped?
• who gets paid here?
• what unlocks the next layer?
• is the bottleneck capacity, qualification, thermal, yield, tooling, or materials?

Also ask:

• if this supplier disappeared tomorrow, how long would the downstream wait for a credible replacement?
• which layer is being paid now versus later?

If needed, read references/question-ladder.md.

If the user wants a graph-style output, load references/graph-schema.md.

Step 3: Hunt the Bottleneck

Now force the real question:

• if AI demand doubles, what breaks first?

Check these bottleneck types:

• physical input shortage
• long lead-time tool / fab capacity
• reliability and qualification delay
• yield bottleneck
• thermal / power limit
• geopolitical or single-region dependence
• single-customer dependency

Rank the bottleneck by:

• concentration
• substitutability
• ramp difficulty
• proof of demand
• whether consensus already sees it

Step 4: Verify With External Evidence

Do not trust one tweet, one chart, or one story.

Use three evidence buckets:

1. company evidence
   • earnings calls
   • investor presentations
   • customer / supplier mentions
   • guidance language
2. industry evidence
   • trade press
   • industry reports
   • capacity / lead-time / deployment news
3. cross-chain evidence
   • multiple companies describing the same stress point from different sides

Preferred sources:

• company IR pages and earnings transcripts
• official filings
• reputable industry reporting
• broker / market research summaries if available

Always label the strongest evidence line and the weakest assumption.

Use this evidence hierarchy:

• strongest: filings, earnings-call transcripts, IR materials, direct customer/supplier disclosures
• strong: official supplier-list changes, design-win announcements, capacity-expansion notices
• medium: reputable industry reporting, broker or market-research summaries
• weak: social posts and unverified forum claims

If multiple companies describe the same constraint from different positions in the chain, say so explicitly. That is higher quality than a single-source story.

If the user wants a 2.0 answer, load references/evidence-ladder.md and tag major claims explicitly as:

• Confirmed
• Inferred
• Weak
• Needs verification

Step 5: Output the Direction First

Default output is a direction, not a stock list.

The first answer should say:

• the next lane worth tracking
• why now
• what confirms it
• what breaks it
• what downstream / upstream companies would feel it first

This is the main output layer.

If the user asks for deeper prioritization, load references/scoring-framework.md and output a lane score.

Step 6: Only Then Offer Candidate Names

If the user pushes deeper, offer 3-7 names.

Split them by role:

• safest executor
• pure bottleneck supplier
• cheaper second-order beneficiary
• early optionality / disruptor

For each name, include:

• role in the stack
• why this name belongs
• what evidence is real
• what still needs confirmation
• main risk

For lower-market-cap names, always include:

• market-cap mismatch versus the layer leader
• current stage: concept / qualification / early ramp / real volume
• why this may still be too early

Do not present them as buy calls.

Step 7: Position-Sizing Logic

If the user asks for weights, use Crux-style discipline:

• leaders / proven executors get more weight
• earlier pre-commercial names get smaller starter positions
• weight increases only if qualification, ramp, and revenue conversion improve

Never size purely off narrative upside.

Output Levels

Choose the shallowest level that satisfies the ask.

Escalation rule:

• first response: direction only
• second response after user follow-up: candidate watchlist
• third response after user picks one lane or one company: underwrite sheet

If the user jumps straight to "give me small caps", first give:

• one-paragraph lane thesis
• one bottleneck summary
• then the names

Do not skip the thesis stage.

Level 1: Directional Lane

Use when the user asks:

• what direction should I study next
• where is the next bottleneck
• what sector is being underpriced

Output:

• thesis in 4-8 short paragraphs
• stack snapshot
• bottleneck call
• proof / disproof checklist

Level 2: Candidate Watchlist

Use when the user follows up with:

• which names are worth watching
• any lower-market-cap ideas
• who are the pure-play beneficiaries

Output:

• 3-7 names
• grouped by role
• one-line thesis, one-line risk, one-line next check

If the evidence base is weak, downgrade the output to:

• names worth validating, not names worth buying

Level 3: Underwrite Sheet

Use when the user asks for one name in depth.

Output:

• why this company matters
• what exact bottleneck it solves
• customer / supplier map
• revenue timing and what to monitor
• failure cases

Read references/output-formats.md for templates.

Recommended Response Skeleton

When answering live user requests, default to this order:

1. thesis sentence
2. stack view
3. bottleneck call
4. evidence and disproof
5. only then names, if requested

This prevents the skill from collapsing into ticker spam.

For L5 lower-market-cap drilldowns, end with two short lines:

• why the market may be underpricing it
• why the user still should not treat it like a proven executor

Tone and Style

Write like a researcher who is trying to catch the market sleeping on a physical constraint.

Good style traits:

• direct
• skeptical
• bottleneck-focused
• willing to say "this is still too early"
• willing to separate "great story" from "real ramp"

Bad style traits:

• generic "AI is bullish" cheerleading
• fake certainty
• too many abstract slogans
• copying public figures' catchphrases

For voice constraints and imitation boundaries, read references/style-and-voice.md.

What Makes a Good Answer

A good answer from this skill:

• starts from the supertrend, not the ticker
• shows the stack clearly
• identifies one primary bottleneck and one backup candidate
• uses at least one external non-social proof source when the user asked for current research
• tells the user what to watch next
• keeps small-cap names as a second-order output, not the starting point

When To Escalate

Ask a clarification only if one of these is genuinely unclear:

• which infrastructure theme the user means
• whether they want direction vs names
• whether current / latest validation matters

If latest validation matters, browse the web and prefer primary sources.