AI Industry Map

AI Core Value Chain, Diffusion Value Chain, Peripheral Value Chain, Module Cycles, Profit Allocation, Competitive Structure, Credit Vulnerability, and Pullback Analysis

Updated Date: 2026-04-22

Chapter 1: Ratings and Valuation Glossary

Supplemental Valuation Glossary

This report also consistently uses the following definitions for valuation status, to avoid conflating "low multiples," "high enthusiasm," and "quality premium" into a single term.

Chapter 2: Executive Summary

This report is lengthy, but only five key points truly drive decision-making. They are sufficient to establish an actionable perspective on the AI industry, with detailed arguments and evidence chains presented in the main body and appendices.

I. Which Three Layers Deserve the Most Attention Now

1. Structural Core Layer（High-Quality Profit Pool + Indispensable + Valuation pullback primarily due to multiple compression rather than credit contraction）——Examples:TSM / KLAC / ASML / NVDA / ANET。
2. Long-Cycle Persistent Delivery Layer（Operational delivery slower than the "hottest" layer, but with higher order and backlog visibility）——Examples:ETN / HUBB / POWL / NVT。
3. High-End Supporting Layer with Importance Exceeding Attention（Supporting components required for AI high-speed boards and advanced packaging and testing）——Examples:Victory Giant Technology / WUS Printed Circuit / Shennan Circuits / TTMI。

II. Three Layers to Avoid Chasing Now

1. High Beta, Highly Crowded Optical Interconnect Layer——LITE / COHR and similar optical modules, where valuation, customer concentration, and generational transition risks are simultaneously amplified.
2. Execution/Volume Ramp Layer——SMCI / DELL / HPE and similar system delivery layers, with thin gross margins, high working capital pressure, and an unstable financial structure.
3. High-Valuation Pure Liquid Cooling and Pure Testing Thematic Stocks——The direction is right, but the price has already run ahead of execution.

III. Top 10 Names to Truly Track Continuously

A condensed list across the three categories, in order:NVDA / TSM / KLAC / ASML / ANET / ETN / NVT / MU（with cyclical discipline）/ Victory Giant Technology / WUS Printed Circuit. These 10 companies cover the three positions: main profit pool, long-cycle undertaking, and key supporting roles. Refuse to compare them with optical module / server / pure testing names on the same scale.

IV. Three Most Critical Triggers

If any of the following signals appears, the entire module master table should be re-ranked:

1. Budget Source Deceleration：MSFT / GOOGL / AMZN / META collectively revise down AI CapEx guidance (magnitude >10%)——Full chain repricing.
2. Physical Supply Constraint Alleviation：TSM's advanced packaging (CoWoS) utilization significantly weakens, or HBM supply release is significantly faster than demand——The multiples of the core profit pool will be compressed first.
3. Deployment Layer Constraints Loosening：The backlog or lead time direction forETN / HUBB / POWL / VRT reverses for the first time——The patience premium for the long-cycle undertaking layer will quickly be given back.

V. Three Most Likely Risks to Disprove Our View

1. General GPU budget significantly diverted to in-house ASICs — NVDA's profit pool is being reallocated, and the market has fully priced in "NVDA's perpetual lead".
2. Power access cycle is longer than AI sentiment cycle + some names have entered the "valuation-business structure mismatch" zone — ETN / NVT, even with sound fundamentals, may see their valuations suppressed during a thematic correction, creating a "right but not patient enough" misalignment.POWL singled out: DC revenue only 2.4%, management CapEx 100% invested in LNG, GM QoQ -3pp, F-C spread at historical high, market pricing at "AI pure beta" 47x P/E is mismatched with its hybrid business structure, has moved from rank 2 to rank 3, rating downgraded to "Cautious Watch (Critical, Trading-oriented)".HUBB singled out: Mgmt-GAAP caliber gap +3-4pp for 3 consecutive years + DMC acquisition's standalone ROIC of only 2.6% dilutes the group + Reverse DCF implies FY26-30 organic CAGR +9.8% as a 1-in-4 tail scenario + Roundtable 5/5 recommends downgrade, retains rank 2 but rating downgraded to "Cautious Watch (Highly Contested)".
3. High-end AI PCB and advanced packaging & testing simultaneously entering supply normalization period — Victory Giant Technology / WUS Printed Circuit / TTMI / FORM's front-loaded valuations are being compressed, the market first downgrades, then slowly confirms profits.

Chapter 3: Reading Guide: How to Use This Report

What Readers Can Take Away from This Report

This report is not adding another "who's relevant" list to the AI industry chain. It aims to address the most common sticking points readers encounter under the AI theme:

• Knowing "who's on the chain," but not knowing who is actually making money.
• Knowing "which layer is most important," but not knowing if it's currently worth paying attention to.
• Knowing "the next round will see a correction," but not knowing which modules will be most vulnerable and which will be more resilient during that correction.

After reading, readers should be able to distinguish three completely different lists:

1. Who is at the center of the main profit pool.
2. Who is the continuous monetization layer in the next phase.
3. Who, while important, is better suited for waiting for a correction or for trading purposes only.

Core Framework: AI Is Not a Chain, But a Closed-Loop Network

The AI industry should no longer be viewed as a horizontal list of "chips → servers → software".
A more accurate understanding is: it is a closed-loop network composed of budget, manufacturing, deployment, and feedback.

This network can be broken down into four layers:

1. Structural Layer — What key elements are in the system.
2. Constraint Layer — What truly limits expansion, profit, and realization.
3. Timing Layer — At what pace each module materializes, and at what pace the market prices it.
4. Valuation Layer — Which are suitable for long-term research, which are better observed after a correction, and which are more akin to trading-oriented cautious watches.

The four layers are sequentially connected, not parallel tags: The Structural Layer answers "What's inside" → The Constraint Layer answers "What's holding it back" → The Timing Layer answers "When will it materialize" → The Valuation Layer answers "Which piece to acquire now".

How to Read This Report

If this is your first time encountering this map, the suggested reading order is:

1. First, read Chapters 5 to 7 to establish why this chain is not a "single line" but rather a three-layer structure of main chain / diffusion chain / peripheral chain.
2. Then, read Chapters 8 to 9, using the module master table and evidence cards to place "structure, constraints, timing, profit, credit, and actions" into a unified working framework.
3. Next, read Chapters 15, 18, 19, 20, 24, 25, 32, and 34, which correspond to the three rankings, cycle clock, order waterfall, profit distribution, competitive structure, credit pressure, valuation framework, and correction actions, respectively.
4. Finally, read Chapter 36 to transform the continuous monitoring list and trigger conditions into a subsequent review worksheet.

If you are already familiar with the AI industry chain, what's most worth looking at first is not the three rankings in Chapter 15 itself, but rather:

• In Chapter 8, how the module master table is defined
• In Chapters 18, 19, 22, 28, and 32, which variables, if changed, should trigger a re-sorting of the entire module master table and the three rankings in Chapter 15

How the Four-Layer System Diagram Corresponds to Each Chapter in the Main Text

• Structural Layer — Full Main Chain Map / Three Rankings / China Chain / Overall Chain Node Table
• Constraint Layer — Which segments warrant continuous attention / Profit Distribution / Customer Concentration and Bargaining Power / Supply Elasticity and Alternative Paths
• Timing Layer — Module Cycle Clock / Order Waterfall and Transmission Lag / Company's Position in the Cycle / Bull / Base / Bear Scenarios
• Valuation Layer — Crowding and Holder Fragility / Tiered Valuation Framework / Deleveraging and Credit Pressure / Correction Action Handbook

Four Core Charts

Chart 1: Budget -> Orders -> Revenue -> Cash Waterfall Chart

Chart 2: Two-Dimensional Compression Chart of Profit Pool Depth × Supply Rigidity

Chart 3: Module Cycle Clock Compression Chart

Chart 4: Crowding × Financial Resilience Quadrant Chart

Chapter 4: What Questions Does This Article Answer?

This article does not assume readers are already familiar with the professional frameworks of semiconductors or data centers.

It only aims to answer a few of the most important questions:

1. Where exactly are the most critical links in the AI industry chain?
2. Which companies are truly on the main chain, which merely benefit adjacently, and which are merely external reflections?
3. Where are the hottest segments in the current market, and which segments are more important than their current market popularity suggests?
4. Among often-overlooked segments like power distribution, power, circuit boards, testing, and packaging, which specific targets are worth examining individually?
5. How should names like FormFactor, Shenghong Technology, ETN/HUBB/POWL/NVT, which are easily relegated to the periphery, be repositioned correctly within the main chain or diffusion chain?
6. If only the most practical list were to be kept, which companies should be included in the AI core node Top 20?
7. What cyclical position are these segments in, and which layers are currently more suitable for research and which for avoidance?
8. How does AI CapEx transmit down the industry chain, which segments receive orders first, and which realize profits later?
9. Within a thematic chain, who captures revenue, who captures gross profit, and who ultimately delivers higher returns to shareholders?

Chapter 5: Conclusion First: AI is Not a Single Line, But a Three-Layer Structure

Drawing the AI industry chain as a single horizontal line will almost certainly lead to misleading judgments.

A more accurate approach is to divide it into three layers:

The Most Critical "Compute and Manufacturing Main Chain"

This layer determines:

• Whether there is sufficient computing power
• Whether there is sufficient high-bandwidth memory
• Whether there are advanced packaging and advanced processes
• Whether there is manufacturing yield and process control

Representatives of this layer are:

• Demand Source Platforms: MSFT / GOOGL / AMZN / META / ORCL
• Accelerators: NVDA / AMD
• HBM / Memory: SK Hynix / Samsung / MU
• Advanced Processes and Packaging: TSM
• Equipment and Process Control: ASML / KLAC / LRCX / AMAT

First-Tier Diffusion Chain

This layer determines:

• Whether large-scale clusters can be interconnected
• Whether racks and servers can be delivered
• Whether power and heat can be managed
• Whether high-speed interconnects and board-level designs can keep pace

Representatives of this layer are:

• Networking: ANET / AVGO / MRVL
• Optical Modules: LITE / COHR / CIEN / AAOI
• Servers and Systems: SMCI / DELL / HPE / Foxconn Industrial Internet / Inspur Information
• Power Distribution and Supply: VRT / ETN / HUBB / POWL / NVT
• Global Electrical Equipment: Schneider Electric / Siemens / ABB
• High-Frequency High-Speed PCBs: Shenghong Technology / WUS Printed Circuit / Shennan Circuits / TTM Technologies

Peripheral Commercialization Layer

This layer determines:

• How AI transforms into enterprise workflows and software revenue
• How data center construction maps to utilities, real estate, and EPC

Representatives of this layer are:

• Software and Workflows: CRM / NOW / PLTR / DDOG
• Utilities and REITs: CEG / VST / NRG / EQIX / DLR

This layer is important but cannot be mixed into the same pool as the physical main chain.
The reason is simple: their methods of benefiting from AI, realization timelines, and valuation methodologies are all different.

Chapter 6: How to Determine if a Segment is "Core"

For readers, there's no need to memorize professional terminology; just remember these three points:

Is it an indispensable critical link?

If this link is absent, the entire chain would struggle to proceed, making it highly likely to belong to the core layer.

For example:

• Top-tier GPUs
• HBM
• Advanced processes and advanced packaging
• EUV Lithography
• Process control and defect inspection

Is it where profits are realized earliest?

Some segments, though important, realize profits very late; others, while not the most prominent, confirm profits quickly.

For example:

• NVDA, TSM, MU are already strong realization layers.
• ETN/HUBB/POWL are more like slow-realization, long-cycle benefit layers.
• CRM/NOW represents a different software migration cycle.

Is its supply difficult to rapidly replenish?

If supply is easily expandable, then it is more like a high-elasticity adjacent layer, rather than a long-term scarce link.

For example:

• Optical modules and some liquid cooling solutions typically have faster supply expansion than HBM and advanced packaging.
• Server systems and high-frequency high-speed boards are important execution layers, but their long-term bargaining power is typically lower than that of core links in the main chain.

Chapter 7: Full Map of the AI Main Chain: Which Segments Belong to Core, Diffusion, Peripheral

Core Main Chain

First-Tier Diffusion Chain

Peripheral Layer

Chapter 8: Module Master Worksheet: Integrating Structure, Constraints, Timing, Profit, Credit, and Actions into a Single Worksheet

The preceding sections addressed:

• How the AI value chain is broadly segmented
• Which companies and modules are categorized as core, diffusion, or peripheral

However, for this report to be useful long-term, mere narration is insufficient.
Each module must also be condensed into a single master template, ensuring that future updates do not revert to discussions about "which sector is the hottest."

Master Worksheet Fixed Fields: 8 Core Fields + 4 Execution Fields

The module master worksheet is fixed at 12 columns.
The first 8 columns are for structural assessment, and the subsequent 4 columns ensure executability.

Master Worksheet Completion Discipline

This section is not a formal requirement, but rather to prevent research from reverting to a "thematic hype table."

First, complete the Asset Node, then fill in other columns.
Because almost all subsequent columns depend on the first column.

For example:

• "Optical modules" are not the ultimate Asset Node; "high-speed optical connectivity capability" is more accurate.
• "Power distribution and electricity" are not the ultimate Asset Node; "data center power-up, distribution, and protection capabilities" are more accurate.
• "Servers" are not the ultimate Asset Node; "AI system delivery and integration capabilities" are more accurate.

Second, write only 1-2 critical constraints for the Constraint Node.
Truly useful constraints are those that alter module ranking, not just a risk checklist.

Third, the Timing Status must include two clocks:

• Operational realization clock
• Market pricing clock

For the same module, operations might still be in the early-to-mid stage, while pricing has already entered the mid-to-late stage crowded zone.
This is precisely the root cause for many AI names where "fundamentals are correct, but the stock price is already too far ahead."

Fourth, Profit Pool Depth must be written separately from importance.
Servers, PCBs, and test chains may all be important, but that does not inherently mean they reside in the deepest profit pools.

Fifth, the Current Rating must be written in actionable research rating language.
The master worksheet is not an opinion sheet; it is a working sheet.
Therefore, 'Deep Focus,' 'Focus,' 'Neutral Focus,' 'Cautious Focus (Trading-oriented),' and 'Cautious Focus (Pending Verification)' are more useful than colloquial judgments like 'long-term bullish,' 'cheap,' or 'expensive.'

Module Master Worksheet Pre-filled for Current AI Industry

The table below is not a one-time opinion sheet, but a working template for future review.
It integrates the three-layer structure, module cycle clock, order waterfall, profit distribution, competitive structure, supply elasticity, financial grouping, crowdedness assessment, and pullback actions discussed later in the text.

How to best use this master table

First, it is not primarily a stock pool, but rather a module sorter.
It solves the question of: which are truly core, which are second-stage beneficiaries, which are important but have shallow profits, and which are more like high-beta trading layers.

Second, do not update all 12 columns simultaneously each time.
The first three columns to check are usually:

• Whether the constraint nodes have changed
• Whether the clock position has changed
• Whether the current valuation status has changed

Because once these three columns change, subsequent action recommendations often change along with them.

Third, what truly triggers a complete table reshuffle is not stock price fluctuations, but changes in structural variables.
The most typical triggers include:

• Platforms collectively revise down AI CapEx
• HBM supply growth outpaces demand
• TSM advanced packaging utilization weakens
• ANET and other network layers confirm cluster expansion is no longer accelerating
• Optical module ASP and orders both weaken
• Backlog or lead time direction for ETN / HUBB / POWL / VRT reverses
• FORM / TER / COHU AI testing demand no longer materializes, but valuations remain high

Lightweight Scoring Version: Enable the Master Table to Support Sorting, without Flattening Research into a Mere Scoring Tool

If this master table needs to be further converted into a sorter, a lightweight scoring version can be used.
The principles here are:

• Scoring is only for auxiliary sorting and does not replace the main text's judgment.
• Do not score all 12 columns; only score the 5 columns that most impact sorting.

The fixed 5 scoring dimensions are:

These 5 items sum to 25 points.
25 points is not the goal itself; it merely helps you quickly distinguish:

• Which modules are worth including in the core research pool
• Which modules are more suitable for the research pool
• Which modules are limited to transactional cautious attention

Illustrative Lightweight Scoring for Current AI Modules

The most important thing here is not the scores themselves, but the two structural facts they reveal:

First, a high score does not mean it should be added to the watch list immediately.
GPU, HBM, advanced packaging, and key equipment score highly because they are closer to true hard constraints and deep profit pools;
however, their current pricing value may not necessarily be high.

Second, the modules most deserving of a "Watch" rating right now are not necessarily those with the highest scores.
Power distribution, protection, and power access form a more balanced combination across structural quality, supply rigidity, financial resilience, and current pricing value.
This is also why it has consistently been given a more prominent position in this report.

Fixed Financial Metric Template: Don't Just Look at P/E in the Future

Standardize financial metrics, dividing them into three layers:

Layer 1: General Financial Metrics for All Modules

Layer 2: Turnover Metrics More Relevant for Manufacturing and Diffusion Chains

Layer 3: Module-Specific Financial Metrics

Three-Part System: All Future Rankings Should Ideally Be Divided into Three Score Sets

To avoid a "one-score-fits-all" approach, this version recommends fixing three sets of scores:

These three sets of scores correspond to three completely different rankings:

• Structural Core Pool Based on Structural Quality Score
• Current Research Priority Pool Based on Current Valuation Score
• Trading / Pullback Pool Based on Financial Resilience Score and Price Position

This is closer to real investment decisions than a single overall ranking.

Chapter 9: Module Evidence Cards: Enabling Every Judgment to Be Traceable and Falsifiable

Truly top-tier long reports are not about having more opinions, but about **every module being auditable and refutable**.

Therefore, we assign a fixed minimum evidence card format to each important module:

How to Use Evidence Cards

For subsequent updates, each module does not require a complete rewrite; only these 6 cells need to be updated first:

1. Whether supporting facts have been added or falsified
2. Whether weakening facts have become stronger
3. Whether the best expression has switched from one name to another
4. Whether the most dangerous misinterpretation has already been made by the market
5. Whether the invalidation trigger is approaching

If these 5 steps are not completed first, any new ranking will not be sufficiently auditable.