Forty years ago, on October 22, 1978, Deng Xiaoping visited Japan for the first time. Traveling the 370-kilometer journey from Tokyo to Kyoto on the world's first high-speed railway—Japan's 'Hikari' Shinkansen—the Japanese accompanying personnel asked him for his impressions. He said: "It just feels like it urges one to run faster. So right now, we are just suited to ride such a train."
AI also has that effect of urging people to run faster.
Over the past two years, Nvidia's revenue has surged from $60 billion to $216 billion, and its stock price has increased tenfold. The wave of investment around AI has swept the globe—optical modules, data centers, cooling, robotics, AI applications—one wave after another. Every day there are new stories of price surges, and every day someone regrets not getting in earlier.
But while AI urges one to run, before running, one must first see the road clearly.
AI is the longest track our generation can encounter. The internet took ten years from 1995 to Google's IPO, and another eight years to Facebook's IPO. In between, it experienced the 2000 bubble burst, with the NASDAQ falling 78%. AI will likely follow a similar path—we might currently be in a position similar to 1998 or 1999. The truly biggest opportunities might only appear after the future bubble bursts, or perhaps they are hidden in some corner nobody is paying attention to today.
Currently, model capabilities are advancing at a rapid pace, capital is pouring in frantically, and valuations are pushed to uneasy heights. In this environment, there are two types of people:
The first type rush in to buy now—gambling that they've timed it right. They might make money, but are more likely to buy halfway up the mountain and then be shaken out by a correction.
The second type wait for the crash—but the problem is, when the crash really comes, will you dare to buy? Do you know what to buy? If you know nothing about this industry, you will only panic more in the face of panic.
I choose a third way: Don't rush to buy stocks now, but first build a position—build a 'knowledge position'.
Because no matter how AI develops, when the real opportunities appear, if we don't want to miss them—we must first become experts with a comprehensive understanding of the entire industry. So-called 'killer intuition' is nothing more than coming from a cognitive state of 'having a clear mental map'.
Starting today, I will begin doing something slow and 'dumb': systematically researching the AI industry from a holistic perspective, studying it bit by bit, understanding the entire AI industry chain from start to finish. Who is making money? Where does the money come from? Where does it flow? Who is irreplaceable? Who is feeding on leftovers?
So that when the day comes that the market gives us an opportunity—whether it's a crash, a correction, or some overlooked corner—I can make a judgment in seconds: 'Is this price worth acting on?'
Furthermore, in doing this, I will have two differentiators:
First, my investment foundation is solid. I have extensive experience and an extremely fast pace of evolution in investing. My return rate over the past three years, as my long-time followers are very clear, has reached a level few can match. Of course, the key isn't the return rate, as that might involve luck. The most important thing, and what is generally recognized, is my pace of evolution—I think this is even more crucial in the AI era. It's not about who is better, but about who evolves faster.
There's no need to dwell on the past. The future starts now. Let's 'wait and see'.
Second, I focus on one thing: how does this thing make money? My rapid evolution in recent years is mainly due to my focus: I only pay attention to the wealth opportunities behind phenomena. Most of the articles we see now teach you to use new Skills, new GitHub repos, pursuing trends and new things every day. These things are important, but from an investor's perspective, I care more about the wealth opportunities behind them.
When the iPhone 4 was released, did you, like others, marvel at the phone's design and performance, or did you research the investment opportunities behind it?
This article is the first in a series of research, aiming to do one main thing: light up the map. If systematically researching the entire AI industry chain is like playing a large open-world game—the first step isn't to rush to fight the Boss, but to first light up the map: which major regions, which key nodes, what is the main quest, what are the side quests. Once the map is clear, no matter what situation arises later, judgments can be made in seconds.
Chapter 1: Why View AI from a Holistic Perspective?
Nvidia's tenfold increase in two years is the most dazzling story in AI investing. But if you only see Nvidia, it's like only seeing one tree—you'll miss the structure of the entire forest beneath it.
Every major technological wave sees money spread outward along the industry chain, layer by layer. This has been repeatedly proven in history:
In the internet era, the first wave of money rushed into Cisco (network equipment), the second wave into Google, Amazon (platforms), the third wave into Facebook, Netflix (applications). In the mobile internet era, the first wave was Qualcomm (chips), the second wave was Apple (terminals), the third wave was WeChat, TikTok (super-apps).
AI is no exception. We can see a rough diffusion chain:
First Circle (2023-2024, already fully priced): GPU—Nvidia
Second Circle (2024-2025, currently being priced): Optical Interconnect, Power—LITE up 16x, Vertiv up 10x
Third Circle (2025-2026, not yet fully priced): Cooling, Storage, Specialized Foundry
Fourth Circle (2026+, awaiting catalyst): AI Applications, Energy Infrastructure, Robotics
For investors, the key insight is: The more foundational the infrastructure layer, the fewer players, the lower the substitutability, and the stronger the pricing power.
There might be thousands of companies competing in the 4th layer AI applications. This is why Nvidia earns $216 billion a year, while most AI application companies are still losing money.
But this also means that within the second, third, and even fourth circles of the infrastructure layer—those companies not yet labeled as 'AI concepts' by the market—may hide a wealth of opportunities. We need to first understand which players exist, what they do, and what they are worth.
Understanding this is significant because: When future market corrections, panic, or divergence occur, we will know where we should be looking.
The diffusion circles described above outline the sequence of market sentiment and capital flow—what money chases first, what later. But to truly understand the business logic of each segment, another map is needed: the hierarchical structure of the industry chain. Next, we will deconstruct it layer by layer, from the bottom up.
I divide the entire AI industry chain into a 4-layer structure, 4 main quest maps.
Chapter 2: Four-Layer Structure, Four Main Quest Maps
The four maps are: Computing Power Infrastructure, Model Layer, Middleware, Application Layer, plus one ultimate constraint: Power.
First Layer: Computing Power Infrastructure—The 'Engine' of AI
This layer is the physical foundation of the entire industry chain. All money—no matter which layer it flows in from—will ultimately settle here.
(1) Chip Design: The Arms King
Nvidia is the undisputed hegemon. In FY 2026 (ending January 2026), total revenue was $216 billion, with data centers contributing $193.7 billion—just two years ago it was less than $50 billion. This growth rate is unprecedented in semiconductor history.
What do these numbers mean? A specific example: training a cutting-edge large model costs hundreds of millions of dollars just for GPUs. And training is a one-time cost; after the model goes live, it needs to process hundreds of millions of user requests daily, each consuming computing power—this is the 'inference' cost. A model's lifetime inference cost can be more than ten times its training cost. This means as long as AI is being used, Nvidia continues to collect a 'tax'.
Nvidia's moat isn't just hardware. Its real barrier is CUDA—a software ecosystem with over 5 million developers. Like iOS for Apple, CUDA makes it hard for users to leave once they're in. AMD (MI300X) and Intel (Gaudi) are catching up, but the ecosystem gap is at least several years.
Another route is custom AI chips. Broadcom provides custom designs for Google's TPU, Amazon's Trainium, etc. The logic is simple: tech giants don't want to be 'choked' by one company forever. But at least for now, self-developed chips are supplements, not replacements.
Core Question: How long can Nvidia's monopoly last? Duan Yongping also said he doesn't understand—"Nvidia will definitely still be around in 10 years, but will it still hold its current market position?" This is a question worth trillions of dollars. And behind this, chip manufacturing involves a long industry chain, which has already boosted many companies. I will pay more attention to this.
(2) Chip Manufacturing, Packaging & Memory: The Armory
Chips designed need to be made. TSMC almost monopolizes the manufacturing of the world's most advanced AI chips. Nvidia, AMD, Broadcom, Apple's core chips are all fabricated by TSMC. In the 3nm, 2nm race, Samsung and Intel's foundry businesses lag far behind.
A more critical bottleneck is High-Bandwidth Memory (HBM). No matter how powerful an AI chip's computing power is, if data can't be 'fed' in, it's useless. SK Hynix leads the HBM field, with HBM3E being almost an exclusive supplier to Nvidia. Samsung and Micron are catching up, with a significant yield gap.
Advanced Packaging (CoWoS) is another capacity bottleneck—supply has been unable to meet demand for over a year.
Core Question: TSMC and SK Hynix's capacity is power. Whoever controls capacity controls the pace of the AI arms race.
(3) Optical Interconnect & Networking: The Nervous System
AI training clusters have expanded from thousands of GPUs to hundreds of thousands. How do chips communicate at high speed? Traditional copper cables hit a physical limit beyond 800Gbps—signal attenuation, power consumption surge, heat dissipation out of control. Optical interconnect is the only way out; this isn't something engineering optimization can solve, it's a hard constraint set by the fundamental laws of electromagnetics.
Key players: Lumentum (LITE, InP laser leader, 16x stock), Coherent (COHR, optical vertical integration), Tower Semiconductor (TSEM, silicon photonics foundry, I've previously written in-depth reports on this), Arista Networks (ANET, AI data center switches), Astera Labs (ALAB, connectivity chips).
Core Question: Optical interconnect is a second-circle opportunity—already being priced, but perhaps not fully priced yet. The key is distinguishing which companies still have room, and which are already priced in. I've recently written several reports related to this.
(4) Cooling & Power Supply: The City Sewer
Nvidia's latest GB200 cabinet power consumption is as high as 120 kilowatts. Putting tens of thousands of cards together generates astonishing heat. Liquid cooling has gone from 'optional' to 'essential'. Microsoft's two-phase immersion cooling technology has already reduced Azure server cooling energy consumption by 95%. Vertiv (VRT) is the leader in this field, with nVent (NVT), Modine (MOD) also growing rapidly.
Core Question: Not sexy, but indispensable. Typical third-circle—most people don't see it, but without it, AI data centers can't run. I will have related reports coming soon.
(5) Servers & Data Centers
Dell, Supermicro integrate chips, memory, networking, and cooling into AI servers. Equinix, Digital Realty provide physical facilities. CoreWeave (IPO expected 2025) is a representative of pure GPU cloud.
(6) Cloud Computing Platforms: Computing Power Wholesalers
AWS, Azure, GCP are the 'wholesalers' of computing power—the three clouds together account for about 65% global market share. Oracle became an unexpected winner with its AI cloud growth.
Second Layer: Models & Tools—The 'Operating System' of AI
This is the most watched, fastest-growing, but most uncertain layer in the AI industry chain.
Five strong contenders: OpenAI (GPT), Anthropic (Claude), Google (Gemini), Meta (Llama open-source), xAI (Grok). The revenue growth in this layer is staggering—Anthropic's ARR (Annualized Recurring Revenue) soared from $1 billion at the end of 2024 to $9 billion by the end of 2025, and surpassed $30 billion by April 2026.
Salesforce took 20 years to reach $30 billion in annual revenue; Anthropic did it in less than 3 years. OpenAI's current ARR is about $24 billion; the two combined exceed $50 billion. Model companies are no longer 'cash-burning stories', but real, gold-earning businesses.
But behind the revenue surge, there's a noteworthy structural change occurring: The focus of AI computing power is shifting from 'training' to 'inference'.
Over the past two years, AI's main computing power consumption was on training large models—pouring massive amounts of data to teach the model to understand the world. But once a model is trained, what follows is 'inference'—actually having the model answer questions and perform tasks.
Research by Deloitte shows that inference's computing power consumption had already surpassed training by the end of 2025, accounting for over 55% of AI cloud infrastructure spending. Some even point out, "In the past, 80% of computing power was spent on training and 20% on inference. In the future, this ratio will reverse."
What does this mean? The inference market may be far larger than the training market (projected to reach $255 billion by 2030), and inference's requirements for chips differ from training—it emphasizes cost efficiency and low latency more than extreme peak computing power. This could be a breakthrough point for challenging Nvidia's monopoly: AMD, Marvell (just received a $2 billion investment from Nvidia), and various self-developed chips are all targeting the inference market.
The most thought-provoking question in this layer is: Will AI models form an oligopoly, or will they be 'commoditized'?
Meta's Llama is free and open-source; DeepSeek created a competitive model at extremely low cost. GLM-5's current API packages are out of stock. Open-source is lowering the barrier to entry for the model layer. But 'commoditization' isn't that simple either—the capability gaps between models are narrowing but haven't disappeared.
Especially in deep usage scenarios, the experiential differences between models remain significant. Moreover, enterprises' API integrations, workflow customizations, and data accumulation create switching costs. The final landscape might be neither 'winner-takes-all' nor 'fully commoditized', but somewhere in between—a few major models occupy the primary market but maintain differentiated competition among themselves.
If profits in the model layer are compressed by open-source, real value will shift upward and downward. Upward to the infrastructure layer because everyone needs to run models, and computing power demand increases rather than decreases. Downward to the application layer because calling costs decrease, making AI applications easier to monetize. This process of profit redistribution might be one of the most important variables in the AI industry chain over the next few years.
Third Layer: Middleware & Platforms—The Glue Layer
The middle layer connecting models and applications. Representative companies: Scale AI (data labeling & AI evaluation, valuation $13.8 billion), LangChain (LLM application development framework), Hugging Face (model sharing platform, the GitHub of AI).
Most companies in this layer are not yet public and are relatively small. But once the AI application layer explodes, these 'glue' companies might experience explosive growth—just like Shopify and Stripe rose with the e-commerce boom. Worth continuous attention.
Fourth Layer: Vertical Applications—The Money Entry Point
Where AI directly creates value for end-users. Several directions:
Enterprise AI Platforms: Palantir sells AI operating systems to governments and enterprises. ServiceNow, Salesforce are grafting AI onto traditional SaaS.
Code Tools: GitHub Copilot is the de facto standard; Cursor is challenging it. The logic is clear—if AI can double programmer efficiency, every enterprise will pay.
Medical AI: Isomorphic Labs (under Alphabet, AlphaFold lineage) might be the most noteworthy long-term prospect, potentially IPO in 2027.
Robotics & Embodied AI: The direction with the largest long-term TAM (Total Addressable Market). Tesla Optimus, Figure AI, Unitree Robotics. But it's still very early.
Autonomous Driving: Waymo has the most mature commercialization; Tesla FSD is catching up with a vision-only approach.
The application layer is where a hundred flowers bloom and also the hardest layer to pick winners. But a noteworthy trend is: The global AI application market size is projected to exceed the upstream infrastructure market for the first time in 2026—money is shifting from 'building the city' to 'opening shops'. Meanwhile, AI Agents (autonomous agents) are becoming a new form of enterprise applications. By the end of 2026, over 40% of enterprise applications are expected to contain built-in AI Agent functionality, compared to less than 5% in 2025.
Cross-Cutting Dimension: Energy—The Ultimate Constraint of AI
All layers cannot avoid one question: Where does the electricity come from?
AI data center power consumption is growing exponentially. Microsoft has $80 billion in Azure orders that cannot be delivered due to insufficient power. This has sparked a wave of energy investment: Constellation Energy (nuclear), NuScale and Oklo (small modular reactors), GE Vernova (gas turbines).
AI will continue to expand; energy infrastructure is a derivative sector with extremely high certainty.
Chapter 4: Four Questions Beyond the Consensus
After drawing the map, the most valuable part isn't confirming consensus, but identifying what the market might be overlooking. Currently, I'm focusing on 4 questions, and subsequent research will start more from these angles.
Question 1: The shift from training to inference—whose fate will it change?
Over the past two years, the main demand for AI computing power was training large models. But now inference (making models actually work) has surpassed training to become a larger market. Inference has different chip requirements than training—more focused on cost-performance ratio than ultimate computing power.
This might open a window: Nvidia's monopoly in the training market is almost unshakeable, but the inference market is more fragmented. AMD, Marvell, Broadcom, and various self-developed chips all have opportunities. Meanwhile, the 'continuous consumption' nature of inference means computing power demand isn't a one-time event but grows continuously with AI application adoption—good news for the entire supply chain.
Question 2: Where is the return on the $600 billion investment?
In 2026, the capital expenditures of the five major tech giants will exceed $600 billion, but the revenue generated by AI applications is roughly a fraction of that figure. A similar input-output gap in history only occurred once—the telecom infrastructure boom in the late 1990s. The outcome then was bankruptcy for many fiber optic companies.
Of course, the key difference is: telecom companies back then relied on debt; today's tech giants rely on their own profits, with debt-to-asset ratios at historical lows. But if AI application monetization speed can't keep up, the capital expenditure growth rate will inevitably slow down—and this will ripple through the entire supply chain. Which companies' risks does this pose?
Question 3: What does the landscape of the second and third circles look like?
Nvidia is the first circle, already fully researched and priced. Optical interconnect and power supply are the second circle, being re-recognized by the market. What about the third circle? Cooling, specialized foundry, AI security, edge inference chips—which companies are in these segments? What are their business models? What is the competitive landscape? If these aren't clarified now, it will be too late when real opportunities appear. This is precisely what the subsequent layer-by-layer research aims to do.
Question 4: How does geopolitics affect the industry chain?
The U.S. export controls on AI chips to China are splitting the global AI industry chain in two. Nvidia's H20 is banned; China is building an independent AI infrastructure set. This means two parallel industry chains are both investing, potentially making the total volume larger than expected. But it also means some suppliers face the risk of 'choosing sides'.
Chapter 5: The Path Forward
The map is drawn; next is the main quest.
I will start from the first layer, delving into each segment one by one. Like clearing areas in a game—first do the main quest (the most core companies and logic of each layer), then the side quests (marginal but potentially surprising corners).
At each stop, clarify three things: What is the business model of this segment? What does the competitive landscape look like? What valuation level is it at? Once these three things are clear, no matter how the market changes in the future, we will have the basis for judgment.
Some Closing Remarks
While writing this industry chain overview, I remembered the LITE story.
I previously did an in-depth review of Lumentum (LITE) on my public account: 'How did others catch LITE's 20x in a year?' It's a textbook case: mid-2024, the market still viewed it as a 'telecom cycle stock', unwanted at $50 per share. But its essence was the 'nervous system' of AI data centers, with a 50-60% global share in InP lasers, the physical limits of copper cables, management expanding capacity counter-cyclically during losses, and book asset value higher than market cap.
All information was public, but I didn't have an industry chain map in my mind to recognize it.
Ultimately, all missed opportunities are not due to 'acting too slowly', but to 'researching too little'.
That's why I want to build a 'knowledge position'. AI is a sufficiently long track—long enough not to need anxiety about not getting on board now, but also not to do nothing and just wait. Understanding every layer, every segment of the industry chain is itself the best preparation. When the day comes that the market gives us an opportunity—whether in the ruins after a bubble burst, or at some suddenly appearing inflection point—with a map in hand, a judgment can be made in seconds.
'Killer intuition is not innate; it's earned through thousands of hours of research.'
