On October 22, 1978, 40 years ago, Deng Xiaoping made his first visit to Japan. From Tokyo to Kyoto, a 370-kilometer journey, he traveled on the world's first high-speed rail—the Japanese Shinkansen "Hikari." When asked by the Japanese accompanying personnel how he felt, he said: "I feel like there's an urge to run, so we are just right to be on this kind of train."
AI also has an urge to run.
Over the past two years, NVIDIA's revenue has surged from $60 billion to $216 billion, and its stock price has increased tenfold. An investment frenzy surrounding AI has swept the globe—optical modules, data centers, cooling systems, robots, AI applications, one wave after another. Every day, there are new stories of gains, and every day, people regret not getting in earlier.
But before running with the urge, you must first see the road clearly.
AI is the longest track our generation will encounter. It took ten years from the birth of the Internet in 1995 to Google's IPO, and another eight years to Facebook's IPO. In between, the Dot-com bubble burst in 2000, with the Nasdaq plummeting by 78%. AI will likely follow a similar path—a position akin to 1998 or 1999 currently, with the real opportunity possibly emerging after a future bubble burst, or it may be hidden in some corner that nobody is paying attention to today.
Currently, model capabilities are advancing rapidly, capital is pouring in frantically, and valuations have been pushed to disturbing heights. In this environment, there are two types of people:
One type rushes in to buy now—betting they have timed it correctly. They might make gains, but it is more likely they buy at the peak, only to be shaken out during a pullback.
The other type waits for the crash to happen before acting—but the question is, when the real crash occurs, will you dare to buy? Do you know what to buy? If you know nothing about this industry, you will only be more panicked in the face of panic.
I choose a third type: not rushing to buy stocks now, but first building a position—building a 'knowledge position'.
Because no matter how AI develops, to ensure we don't miss the real opportunity when it arises—we must first become experts who understand the entire industry. The so-called 'killer instinct' is nothing more than cognitive readiness that comes from 'a deep understanding in the heart'.
Starting today, I will embark on a slow and silly task, conducting a top-down analysis of the AI industry, unraveling the AI value chain piece by piece. I aim to understand who is making money, where the money is coming from, where it is going, who is indispensable, and who is merely picking up the crumbs.
When the market presents us with an opportunity—be it a crash, a pullback, or some overlooked corner—I will be able to make a quick assessment in seconds: "Is this price worth getting in?"
Furthermore, there are two differentiators in my approach:
First, I have a solid investment foundation. I possess extensive experience and rapid evolutionary speed in investments. My performance over the past three years, as my long-time followers are well aware, has been exceptional and few have matched my level. However, the key is not the return rate, as luck may play a role, but what is universally acknowledged is my evolutionary speed—especially in the age of AI, it is not about being better than others, but evolving faster than others.
We need not dwell on the past; the future starts from now, and let's "wait and see."
Second, I focus on one thing: How does this thing make money? The rapid evolution I have undergone in recent years is mainly due to my focus: I only pay attention to the wealth opportunities behind the phenomena. Most of the articles you see now teach you new skills, new GitHub projects, and chase bestsellers and new things every day. While these are important, from an investor's perspective, I am more concerned about the wealth opportunities behind them.
When the iPhone 4 was released, did you join others in marveling at the design and performance of the phone, or did you research the investment opportunity behind it?
This article is the first in a series of studies and aims to do one thing: Illuminate the map. If likening the systematic study of the entire AI value chain to playing a large-scale open-world game, the first step is not to rush to fight the Boss, but to illuminate the map first: which are the major zones, what are the key nodes, what is the main quest, and what are the side quests. Once the map is clear, no matter what happens next, a quick judgment can be made in seconds.
Chapter One: Why Look at AI from a Global Perspective?
NVIDIA's tenfold growth in two years is the most dazzling tale in AI investment. However, if you only see NVIDIA, it's like only seeing one tree—you would overlook the structure of the entire forest beneath its feet.
With every major technological wave, money spreads layer by layer along the industry chain. This has been repeatedly validated throughout history:
In the Internet era, the first wave of money flowed into Cisco (network equipment), the second wave into Google, Amazon (platforms), and the third wave into Facebook, Netflix (applications). In the mobile Internet era, the first wave was Qualcomm (chips), the second wave was Apple (devices), and the third wave was WeChat, TikTok (super apps).
AI is no exception. We can see a rough diffusion chain:
First Round (2023-2024, already fully priced in): GPU—NVIDIA Second Round (2024-2025, currently pricing in): Optical Interconnect, Power—LITE up 16x, Vertiv up 10x Third Round (2025-2026, not fully priced in yet): Cooling, Storage, Specialized Manufacturing Fourth Round (2026+, waiting for catalyst): AI applications, Energy Infrastructure, Robotics For investors, the key insight is: The deeper the infrastructure layer, the fewer players, the lower substitutability, the stronger pricing power.
AI application companies at the fourth layer may have thousands of competitors. That's why NVIDIA makes $216 billion in revenue per year, while most AI application companies are still losing money.
But this also means that in the infrastructure layers of the second, third, or even fourth round—those companies that have not yet been labeled with the "AI concept" in the market—there may be a wealth of opportunities. We need to first understand which players are involved, what they are doing, and how much they are worth.
Understanding its significance means: When the future market experiences a pullback, panic, or differentiation, we will know where to look.
The four-layer diffusion circle described above depicts the transmission order of market sentiment and funds—what money pursues first and then next. But to truly understand the business logic of each link, another chart is needed: The hierarchical structure of the industry chain. Next, we will break it down layer by layer from the bottom to the top.
I divided the entire AI industry chain into a 4-layer structure, with 4 main task maps.
Chapter 2: Four-Layer Structure, Four Main Task Maps
The four maps are: Computing Power Infrastructure, Model Layer, Middleware, Application Layer, and there is also an ultimate constraint: Power Supply.
Layer One: Compute Infrastructure - AI's "Engine"
This layer is the physical foundation of the entire industry chain. All money - no matter which layer it flows from - will eventually settle here.
(1) Chip Design: King of Arms
NVIDIA is the undisputed leader. In the 2026 fiscal year (ending January 2026), total revenue was $216 billion, with the data center contributing $193.7 billion - just under $50 billion just two years ago. This growth rate is unprecedented in semiconductor history.
What do these numbers mean? For a specific example: training a cutting-edge large model, the GPU cost alone is in the hundreds of millions of dollars. And training is just a one-time thing; once the model is online, it has to handle hundreds of millions of user requests every day, each consuming compute power - this is the "inference" cost. The lifetime inference cost of a model may be more than ten times the training cost. This means that as long as AI is still in use, NVIDIA will continue to collect "taxes."
NVIDIA's moat is not just hardware. Its true barrier is CUDA - a software ecosystem with over 5 million developers. Like iOS to Apple, once users enter CUDA, it's hard to leave. AMD (MI300X) and Intel (Gaudi) are catching up, but the ecosystem gap is still several years behind.
Another approach is custom AI chips. Broadcom provides custom designs for Google's TPU, Amazon's Trainium, and others. The logic is simple: tech giants don't want to be forever "blocked" by one company. However, at least for now, self-developed chips are a supplement, not a replacement.
Key Issue: How long can NVIDIA's monopoly last? Zhang Yiming once said that he couldn't understand either - "Will NVIDIA still have its current market position in 10 years?" This is a question worth trillions of dollars. And behind this, there is a whole chain of chip manufacturing, which has already boosted many industries. I will pay more attention to this.
(2) Chip Manufacturing, Packaging, and Storage: Arsenal
Designed chips need to be manufactured. TSMC almost monopolizes the manufacturing of the world's most advanced AI chips. NVIDIA, AMD, Broadcom, and Apple's core chips are all fabricated by TSMC. In the race for 3nm and 2nm, Samsung and Intel's fabrication businesses are far behind.
The more critical bottleneck is High Bandwidth Memory (HBM). No matter how powerful the AI chip is, it's useless if the data can't be "fed in." SK Hynix is far ahead in the HBM field, and HBM3E is almost NVIDIA's exclusive supplier. Samsung and Micron are catching up, but the yield gap is significant.
Chip-on-Wafer-on-Substrate (CoWoS) is another bottleneck in the supply chain — demand has outstripped supply for over a year now.
Key Issue: TSMC and SK Hynix's capacity is power. Whoever controls capacity controls the pace of the AI arms race.
(3) Optical Interconnect & Network: Neural System
AI training clusters have scaled from a few thousand GPUs to tens of thousands. How do chips communicate at high speed? Traditional copper wires have hit a physical limit above 800Gbps — signal attenuation, power surge, and heat dissipation issues. Optical interconnect is the only way out, not something engineering optimization can solve, but a hard constraint set by the basic 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 also written in-depth research reports on), Arista Networks (ANET, AI data center switch), Astera Labs (ALAB, connectivity chip).
Key Issue: Optical interconnect is a second-round opportunity — it has already started to be priced in, but may not have been fully priced in. The key is to identify which companies still have room to run, and which ones have already priced in, as I have discussed in multiple recent research reports.
(4) Cooling and Power Supply: Urban Sewer System
NVIDIA's latest GB200 cabinet has a power consumption of up to 120 kW. When tens of thousands of cards are put together, the heat is staggering. Liquid cooling has evolved from being "optional" to "mandatory." Microsoft's two-phase immersion cooling technology has already reduced the cooling energy consumption of Azure servers by 95%. Vertiv (VRT) leads in this field, while nVent (NVT) and Modine (MOD) are also experiencing rapid growth.
Key Issue: Not glamorous, but indispensable. A typical third round — most people can't see it, but without it, AI data centers cannot operate. I will have relevant research reports on this topic coming soon.
(5) Servers and Data Centers
Dell, Supermicro integrate chips, memory, network, and cooling into AI servers. Equinix, Digital Realty provide physical data centers. CoreWeave (IPO in 2025) is a representative of a pure GPU cloud.
（6）Cloud Computing Platform: Computing Power Wholesaler
AWS, Azure, and GCP are the "wholesalers" of computing power — the three major clouds collectively hold approximately 65% of the global market share. Oracle has become an unexpected winner in AI cloud growth.
Layer Two: Models and Tools — AI's "Operating System"
This is the most attention-grabbing, astonishingly fast-growing, yet most uncertain layer in the AI industry chain.
Top Five Battle: OpenAI (GPT), Anthropic (Claude), Google (Gemini), Meta (Llama open-source), xAI (Grok). The revenue growth in this layer is staggering — Anthropic's Annual Recurring Revenue (ARR) soared from $1 billion at the end of 2024 to $9 billion at the end of 2025, surpassing $30 billion in April 2026.
It took Salesforce 20 years to reach $30 billion in annual revenue, while Anthropic did it in less than 3 years. OpenAI currently has an ARR of around $24 billion, with the two together exceeding $50 billion. Model companies are no longer a "burning money story" but a real gold rush business.
Behind the skyrocketing revenue, there is a noteworthy structural shift taking place: The center of AI computing power is shifting from "training" to "inference."
Over the past two years, the primary AI compute consumption has been in training large models — feeding massive amounts of data to help the model understand the world. However, once the model is trained, the next step is "inference" — making the model actually answer questions and perform tasks.
Deloitte's research shows that by the end of 2025, inference compute has exceeded training, accounting for over 55% of AI cloud infrastructure spending. Some even point out that "in the past, 80% of compute was spent on training, and 20% on inference; this ratio will reverse in the future."
What does this mean? The inference market may be far larger than the training market (expected to reach $250 billion by 2030), and the demand for chips in inference is different from training — focusing more on cost efficiency and low latency rather than peak compute power. This could be a breakthrough in challenging NVIDIA's dominance: AMD, Marvell (which just received a $2 billion investment from NVIDIA), and various in-house chip designs are all targeting the inference market.
The most thought-provoking question at this level is: Will AI models form an oligopoly or be "commoditized"?
Meta's Llama is freely available, and DeepSeek has produced a competitive model at a very low cost. The current API package for GLM-5 is currently out of stock, and open source is lowering the barrier to the model layer. However, "commoditization" is not that simple—the gap in capabilities between various models is narrowing but has not disappeared.
Especially in deep-use scenarios, the experience gap between models is still significant, and enterprise API integration, workflow customization, and data accumulation will all incur switching costs. The eventual landscape may not be "winner takes all" or "fully commoditized," but something in between—where a few models dominate the main market but maintain differentiation.
If the profit at the model layer is compressed by open source, the real value will shift to the upper and lower layers. It will shift to the upper layer, the infrastructure layer, because everyone needs to run models, and the demand for computing power is increasing rather than decreasing. It will also shift down to the application layer because the cost of invocation is decreasing, making AI applications more profitable. This redistribution of profits may be one of the most important variables in the AI industry chain in the coming years.
Layer Three: Middleware and Platforms—Glue Layer
The intermediate layer that connects models and applications. Representative companies: Scale AI (data labeling and AI evaluation, valued at $13.8 billion), LangChain (LLM application development framework), Hugging Face (model-sharing platform, the GitHub of the AI field).
Most companies at this level are not yet public and are relatively small in scale. However, once the AI application layer erupts, these "glue" companies may experience explosive growth—similar to the rise of Shopify and Stripe during the e-commerce boom. It is worth ongoing attention.
Layer Four: Vertical Applications—The Money Entry Point
Where AI directly creates value for end users. Several directions:
Enterprise AI Platforms: Palantir sells its AI operating system to governments and enterprises. ServiceNow and Salesforce are integrating AI into traditional SaaS.
Code Tools: GitHub Copilot has become the de facto standard, with Cursor as a challenger. The logic is clear—if AI can double the efficiency of programmers, every company will pay.
Medical AI: Isomorphic Labs (a subsidiary of Alphabet with AlphaFold lineage) may be the most anticipated target in the long term, with a potential IPO in 2027.
Robotics and Embodied Intelligence: The direction with the largest long-term TAM. Tesla Optimus, Figure AI, Ubtech. However, it is currently in a very early stage.
Autonomous Driving: Waymo is the most mature in commercialization, while Tesla's FSD is catching up with a vision-based solution.
The application layer is where there is a plethora of options and also the most challenging to pick winners. But one notable trend is that by 2026, the global AI application market size is expected to surpass, for the first time, the upstream infrastructure market – money is shifting from "building the city" to "opening shops". At the same time, AI Agents are becoming a new form of enterprise application, with more than 40% of enterprise applications expected to have built-in AI Agent functionality by the end of 2026, a percentage that was 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?
The power consumption of AI data centers is growing exponentially. Microsoft has an $80 billion Azure order that cannot be fulfilled due to power shortages. This has led to a wave of energy investment: Constellation Energy (nuclear power), NuScale and Oklo (small modular nuclear reactors), GE Vernova (gas turbines).
AI will continue to expand, and the energy infrastructure is a highly deterministic derivative track.
Chapter Four: Four Issues Beyond Consensus
After mapping out the territory, the most valuable thing is not to confirm consensus but to identify what the market may be overlooking. Currently, I am focusing on four issues, and future research will start more from these perspectives.
Issue One: The shift from Training to Inference – Whose Fate Will Change?
Over the past two years, the primary demand for AI compute power has been training large models. But now, inference (making the model actually work) has surpassed training to become a larger market. Inference has different chip requirements from training – it focuses more on cost-effectiveness rather than ultimate computing power.
This may open a window: NVIDIA's monopoly in the training market is almost unshakable, but the inference market is more diversified, with opportunities for AMD, Marvell, Broadcom, and in-house chip designs. Additionally, the continuous demand nature of inference means that the need for computing power is not one-time but continues to grow with the widespread adoption of AI applications—this is good news for the entire supply chain.
Question 2: Where is the Return on the $600 Billion Investment?
By 2026, the capital expenditures of the five tech giants will exceed $600 billion, but the revenue generated by AI applications is only a fraction of this figure. A similar investment-output gap has only occurred once in history—the late 1990s telecom infrastructure era, which ended with the bankruptcy of numerous fiber optic companies.
Of course, the key difference is: telecom companies relied on debt back then, while today's tech giants rely on their own profits, with historically low debt ratios. However, if the monetization speed of AI applications does not keep up, the growth rate of capital expenditures will certainly slow down—and this will transmit risks to the entire supply chain. Which companies will be at risk?
Question 3: What Does the Landscape of the Second and Third Circles Look Like?
NVIDIA represents the first circle, which has already been thoroughly researched and priced. Photonics and power are the second circle, being reevaluated by the market. So, what about the third circle? Cooling systems, specialty manufacturing, AI security, edge inference chips—what companies operate in these areas? What are their business models? What is the competitive landscape? If these questions are not clarified now, it will be too late when real opportunities emerge. This is what needs to be addressed in the next in-depth research.
Question 4: How Does Geopolitics Affect the Supply Chain?
The U.S. export control on AI chips to China is splitting the global AI supply chain in two. NVIDIA's H20 is banned, and China is building an independent AI infrastructure. This means that two parallel supply chains are both being invested in, possibly exceeding initial expectations. But it also means that some suppliers face the risk of "picking sides."
Chapter 5: What's Next
The map is drawn, and now it's time to follow the main quest.
Starting from the first layer, I will delve into each link step by step. Just like playing a game and clearing one area at a time—following the main quest (the core companies and logic of each layer) and then exploring side quests (the edges but potentially surprising corners).
At each step, figure out three things: What is the business model of this segment? What is the competitive landscape? At what valuation level is it at? By understanding these three things, no matter how the future market changes, we will have a basis for judgment.
A Few Words
As I wrote this overview of the industry chain, I recalled the story of LITE.
I previously conducted an in-depth analysis of Lumentum (LITE) on my WeChat Official Account. Lumentum's stock price increased 20 times in a year. How did others seize the opportunity? This is a textbook case: In mid-2024, the market still saw it as a "telecom cyclical stock" and no one wanted it at $50 per share. However, its essence is the "neural system" of AI data centers, with 50-60% global market share of InP lasers, reaching the physical limit of copper cables, management expanding production countercyclically during losses, and book value exceeding market capitalization.
All information was publicly available, but I did not have an industry chain map in my mind to identify it.
All missed opportunities ultimately stem from "insufficient research" rather than "acting too slowly."
This is why I want to build a "Knowledge Repository." AI is a long enough track—long enough that there is no need to worry about not getting on board now, but also cannot afford to do nothing and just wait. Understanding every layer and every segment of the industry chain is the best preparation in itself. When the market gives us the opportunity one day—whether it is the ruins after the bubble bursts or a sudden turning point—having a map in hand, we can make a judgment in seconds.
"Killer-like intuition is not innate; it is earned through thousands of hours of research."
Welcome to join the official BlockBeats community:
Telegram Subscription Group: https://t.me/theblockbeats
Telegram Discussion Group: https://t.me/BlockBeats_App
Official Twitter Account: https://twitter.com/BlockBeatsAsia
