20-minute guide · 11 chapters

Learn the AI Chain

From sand to chip to data center to your phone. Understand every link in the AI chain, who makes money at each step, and where the future is heading. Read from top to bottom — no prior knowledge needed.

EVERY INDUSTRY · EVERY COUNTRY · EVERY DAY🌍AI ECONOMY🧠ModelsChips🏢Data Centers🔌Power🔗Networking💻AppsEach layer depends on the others. Break one, break all.
20
Minutes to read
11
Chapters
50+
Stock tickers
🤔
Chapter 1

What is AI, really?

Let's start simple. AI is a computer program that learned from looking at TONS of examples.

Imagine you showed a kid 10 million pictures of cats and dogs. Eventually, they would get REALLY good at telling them apart. That's basically what AI does — except instead of cats and dogs, modern AI looks at billions of pages of text, images, code, and conversations.

Once the AI has "studied" all this, it can do amazing things:

✍️
Write essays
🎨
Make pictures
💻
Write code
🤖
Answer questions
⚙️

The two big jobs for AI computers

There are only two main things AI computers do:

  1. 1. Training — Teaching the AI by showing it examples. This is HARD and needs huge computers running for months.
  2. 2. Inference — Actually using the AI to answer your questions. Each answer is fast, but with billions of people asking, it adds up!
Infographic

📊 Training vs Inference: Compared

📚
TRAINING
Teaching the AI
Months
to complete
$1B+
cost per run
Once
per model
HUGE
GPU cluster
💬
INFERENCE
Using the AI
Seconds
per answer
Pennies
per question
Billions
of times/day
10-100x
bigger market
💡Key Takeaway
AI is a program trained on huge amounts of data. It needs special computers for two jobs: training (teaching it) and inference (using it). Inference is going to be 10-100x bigger than training over time — because every person uses AI many times a day.
🧠
Chapter 2

The Brain — AI Models

Models are the actual AI programs. Some are free, some cost billions to build.

Every AI you've heard of — ChatGPT, Claude, Gemini, Llama — is what we call a "model." Think of a model as a brain that was taught for months by reading basically the entire internet.

🔮
GPT-5
OpenAI
🎭
Claude Opus 4.6
Anthropic
💎
Gemini 2
Google
🦙
Llama 4
Meta (FREE!)
🔐

Closed Models (Paid)

  • OpenAI, Anthropic, Google keep their models secret
  • You pay every time you use them
  • Very expensive to build (billions of $)
  • Usually the smartest AI available
🔓

Open Models (Free)

  • Meta gives away Llama for free
  • You can download and run them yourself
  • Getting close to paid models in quality
  • Makes AI cheaper for everyone
💰

How do AI companies make money?

Think of it like electricity. You pay for each "use" (each question, each image, each line of code). Businesses pay billions of dollars per year using AI models through things called "APIs" (fancy word for asking AI questions by code). Microsoft puts Copilot in Office for 400+ million people. Salesforce puts AI in Agentforce for business sales teams. Every question = money for the model maker.

📈 Listed companies you can actually invest in:

MSFT
Microsoft
Biggest AI distributor. Owns part of OpenAI. Copilot in Office everywhere.
GOOGL
Alphabet (Google)
Makes Gemini models. Huge AI compute via Google Cloud + custom TPU chips.
META
Meta Platforms
Makes Llama models FREE. Huge GPU fleet. AI makes ads way better.
ORCL
Oracle
Cloud provider. Hosts big AI workloads. Big partner with OpenAI.
PLTR
Palantir
Sells an AI platform called AIP to governments and big companies.
💡Key Takeaway
AI models are the brains. The smartest ones are getting better fast. But as free models catch up, the companies that distribute AI (Microsoft, Google) and the companies that run AI on their servers become even more important.
Chapter 3

The Engine — GPUs

GPUs are the special chips that make AI possible. Without them, nothing works.

A GPU (Graphics Processing Unit) is like the engine in a race car — super fast and specialized. Originally they were built for video games. Then scientists discovered they were ALSO perfect for AI. Now every AI in the world runs on GPUs.

Infographic

🎯 Who controls the AI chip market?

AI ChipMarket
NVIDIA80%
AMD10%
Custom (Google TPU, etc)7%
Others3%
NVIDIA's CUDA software ecosystem makes switching costs enormous.

🏆 The GPU Leaderboard (2026)

🥇
NVIDIA Blackwell (B200/GB300)
~80% market share
🥈
AMD MI350
The main challenger
🥉
Google TPU v6
Made for Google's own AI
🎯
NVIDIA Rubin (late 2026)
3x faster than Blackwell
👑

Why NVIDIA dominates

It's not just that NVIDIA makes the fastest chip. They also built a software system called CUDA that basically every AI tool in the world uses. Switching away from NVIDIA is like asking everyone to learn a new language. That's their superpower.
$40K+
Price of a single H100 GPU
Some secondary market
80%
NVIDIA's share of AI chips
Huge dominance
$650B+
Cloud giants' AI spending in 2026
Microsoft + Google + Meta + Amazon

📈 The GPU stocks to know:

NVDA
NVIDIA
THE king of AI chips. If you invest in AI, you probably own this.
AMD
Advanced Micro Devices
Main challenger to NVIDIA with MI350 GPUs. Also makes AI-ready CPUs.
AVGO
Broadcom
Makes custom AI chips for Google and Meta. Secretly huge in AI.
MRVL
Marvell Technology
Custom AI chips for Amazon. Also optical connections.
INTC
Intel
Trying to catch up with Gaudi AI chips and chip-making services.
CBRS
Cerebras (NEW IPO!)
Makes wafer-scale chips (one huge chip vs many small ones). IPO Q2 2026.
💡Key Takeaway
GPUs are the picks and shovels of the AI gold rush. NVIDIA (NVDA) is the clear leader with ~80% share. Every dollar spent on AI compute flows through a GPU. When you hear about cloud companies spending "$650 billion on AI" — most of that money ends up buying chips from NVIDIA.
🏭
Chapter 4

The Factory — How Chips Are Made

Here's the wild part: NVIDIA doesn't actually MAKE their own chips. They just design them.

This is the coolest and most important part of the AI chain — and most people don't know about it! Making an AI chip requires a supply chain with just a handful of companies at each step. If ANY one breaks, the whole thing stops.

🔧 Here's how an AI chip gets made:

⚙️
Design
NVIDIA, AMD
🔬
Laser Machine
ASML
🧪
Tools
AMAT, LRCX
🏭
Factory
TSMC
💾
Memory
SK Hynix
📦
Package
TSMC CoWoS
🚚
Ship
To customer
Infographic

🌍 The AI supply chain crosses the world

ASMLTSMCSK HynixNVIDIACloud Giants
Netherlands 🇳🇱
ASML
Makes laser machines
Taiwan 🇹🇼
TSMC
Bakes the chips
Korea 🇰🇷
SK Hynix
HBM memory
USA 🇺🇸
NVIDIA
Designs chips
USA 🇺🇸
Cloud Giants
Buy & deploy
🔬

The most important company you've never heard of

ASML is a Dutch company that makes one thing: the laser machines used to print patterns on chips. Here's the crazy part — they're the ONLY company in the world that can make the most advanced version of this machine (called EUV).

Each machine costs over $200 million and takes 6 months to build. Without ASML, there is no TSMC. Without TSMC, there is no NVIDIA. Without NVIDIA... no ChatGPT, no Claude, no AI.

90%
Advanced chips made by TSMC
Basically a monopoly
$200M
Price of one ASML EUV machine
Can only make a few per year
3
Companies making HBM memory
SK Hynix, Samsung, Micron
📦

The hidden bottleneck: Packaging

Here's something crazy. Modern AI chips aren't ONE chip — they're several chips glued together onto one package with special memory stacked on top. This "packaging" step is called CoWoS and TSMC is basically the only company doing it well. This is the slowest part of the whole supply chain — the REAL bottleneck!

📈 Supply chain stocks (the "picks and shovels"):

TSM
TSMC (Taiwan Semiconductor)
The factory. Makes ~90% of advanced AI chips. Everyone depends on them.
ASML
ASML Holding
Makes the laser machines to print chips. A pure monopoly on the most advanced version.
AMAT
Applied Materials
Builds the other tools inside the chip factories. Every fab uses these.
LRCX
Lam Research
Makes etching and deposition tools. Critical for chip manufacturing.
KLAC
KLA Corp
Inspects chips for defects. Without KLA, yields are terrible.
000660.KS
SK Hynix
Makes HBM memory — the special stacked memory that sits on every AI chip. Market leader.
MU
Micron Technology
American HBM memory maker. Third player in the HBM race.
SNPS
Synopsys
Sells the software tools used to design every AI chip. Quiet but huge.
CDNS
Cadence Design Systems
Another chip design software company. Duopoly with Synopsys.
💡Key Takeaway
The chip supply chain is like a relay race. Each runner is critical — if ANY one drops the baton, the whole race stops. TSM bakes the chips.ASML makes the laser machines. SK Hynix makes the memory. TSMC, ASML, and SK Hynix are the three most important companies in AI hardware — but they're all non-US companies, which is a geopolitical thing worth watching.
🏢
Chapter 5

The Home — Data Centers

Thousands of GPUs need a place to live. Enter: massive computer warehouses.

Picture a warehouse the size of several football fields, filled with thousands of AI computers running 24/7. That's a data center. AI data centers use 5-10x more power than old-style data centers and produce so much heat that liquid cooling (like a water-cooled engine) is becoming standard.

Infographic

🏗️ Inside an AI data center

AI DATA CENTERLiquid coolingGPU racks (120kW each!)500 MW(a small city!)
🖥️
GPU Racks
120kW each
💧
Liquid Cooling
Mandatory now
Power
50-500+ MW
🔗
Fiber Cables
800G+ speed
🚧

The Great Data Center Stall

Here's where it gets wild. Companies want to build SO many AI data centers in 2026 that nearly half of them are delayed or canceled. Why? Not enough electrical transformers, not enough grid capacity, not enough power. The physical world can't keep up with AI's appetite.
10 GW
AI power load by end 2026
Like 10 nuclear plants
5-10 yrs
Grid connection wait times
In key US markets
3-5 yrs
Wait for grid transformers
Now the #1 bottleneck

📈 Data center & infrastructure stocks:

EQIX
Equinix
World's largest data center company. 260+ facilities globally.
DLR
Digital Realty
Huge data center operator focused on wholesale AI campuses.
VRT
Vertiv Holdings
The cooling and power kings of AI data centers. $15B backlog!
ETN
Eaton Corporation
Makes the electrical gear — switchgear, UPS, transformers.
APH
Amphenol
Makes the connectors inside every AI server. Quiet but massive.
💡Key Takeaway
Data centers are the factories of AI. Demand is INSANE but physical reality is the limit.VRT (Vertiv) is a pure-play on AI cooling/power with a $15 billion backlog.EQIX and DLR are "landlord" REITs that own the buildings.
🔌
Chapter 6

The Energy — Power & Nuclear

Data centers are electricity monsters. This turned old boring power companies into AI stocks!

Here's a story nobody saw coming. For decades, US electricity demand was basically flat. Then AI hit. Suddenly data centers need MASSIVE power. And the best way to deliver reliable, 24/7, clean power is... nuclear. Nuclear stocks went from forgotten to one of the hottest AI trades.

☢️

Meta's $10 Billion Nuclear Bet

In 2025, Meta signed deals with CEG, VST,OKLO and TerraPower for 6.6 gigawatts of nuclear power — enough to power about 5 million homes. This was the largest corporate nuclear commitment in American history. All just to power AI data centers!

⚡ The AI power chain:

☢️
Nuclear
CEG, VST
🏭
Power plants
Make electricity
📡
Transmission
Big power lines
🔲
Transformers
GEV — 3-5 yr wait!
⚙️
Switchgear
ETN, VRT
🏢
Data center
AI runs!
Infographic

⚡ How electricity flows from plant to AI

⚠️ BOTTLENECK☢️Nuclear PlantCEG, VSTTransmissionHigh voltage🔲TransformerGEV (3-5 yr wait!)🏢Data CenterAI runs here
⚠️

The transformer problem

Grid transformers are those giant metal boxes you see at substations. They weigh up to 400 tons, cost millions each, and take 3-5 YEARS to build. There aren't enough factories making them, and demand from AI + renewables + grid upgrades means they're the biggest choke point in the whole AI build-out.

📈 Power & energy stocks:

CEG
Constellation Energy
Largest US nuclear fleet. $16.4B Calpine deal. Meta's main nuclear partner.
VST
Vistra Corp
Nuclear + gas power. Signed 2.1 GW deal with Meta.
TLN
Talen Energy
Sold a data center campus to AWS connected directly to a nuclear plant!
GEV
GE Vernova
Makes grid transformers + gas turbines. Multi-year backlog.
PWR
Quanta Services
Builds the transmission lines and electrical infrastructure itself.
OKLO
Oklo Inc
Small nuclear reactors (SMRs). Sam Altman backed. Meta signed a deal!
SMR
NuScale Power
Another SMR developer. Targeting AI data centers specifically.
NEE
NextEra Energy
Biggest renewable energy producer. Solar/wind for data centers.
💡Key Takeaway
Power is the #1 bottleneck of AI. That's turned boring utilities into exciting AI plays.CEG and VST are the main nuclear winners.GEV makes the transformers that are so scarce. OKLO is a moonshot bet on small modular reactors. No power = no AI. Simple as that.
🔗
Chapter 7

The Plumbing — Networking

100,000 GPUs need to talk to each other at lightning speed. That's a lot of plumbing!

Imagine 100,000 kids in a classroom who ALL need to pass notes to each other at the same time. That's basically what AI training does with GPUs. You need super-fast switches and cables (using light beams) to connect them all. This is worth tens of billions of dollars.

400G
Old speed (2023)
gigabits per second
800G
Today's speed (2026)
2x as fast
1.6T
Coming soon
4x as fast!
Infographic

🕸️ How GPUs talk to each other

SWITCHANET 800G24 GPUs passing data 800 billion times per second
🏆

Arista is the quiet AI winner

ANET (Arista Networks) makes the Ethernet switches that sit inside every big AI cluster. They hit $9 billion in revenue in 2025 (up 29%!) and raised their 2026 AI networking target to $3.25 billion. While everyone talks about NVIDIA, Arista is silently eating.

📈 Networking stocks:

ANET
Arista Networks
The king of AI Ethernet switches. Goal: $3.25B AI networking revenue in 2026.
AVGO
Broadcom
Networking silicon (inside every switch) + custom AI chips. Hidden giant.
COHR
Coherent Corp
Makes the 800G/1.6T optical transceivers (the light-based connectors).
LITE
Lumentum Holdings
Photonic chips and lasers for data center optical networking.
CIEN
Ciena
Long-distance optical for connecting data centers across cities.
MRVL
Marvell Technology
Custom chips + photonic interconnects (acquired Celestial AI).
💡Key Takeaway
Networking grows proportionally with GPU count. More GPUs = more switches needed = more optical cables needed.ANET is the cleanest pure-play on AI networking.AVGO is more diversified (also makes custom AI chips).
💻
Chapter 8

The Apps — Where AI Meets You

All this hardware exists to power apps real people use. Here's where AI becomes useful.

Okay, so we've got brains (models), engines (GPUs), factories (foundries), homes (data centers), energy (power), and plumbing (networking). What's it all FOR? Apps that help people and businesses!

💼 Where AI is actually used today:

👨‍💻

Coding Tools

Claude Code, Cursor, GitHub Copilot — writes code for programmers

MSFTGOOGL
🏢

Enterprise AI

Copilot in Office, Agentforce in Salesforce — every business wants this

MSFTCRMNOW
💬

Chatbots

ChatGPT, Claude, Gemini — billions of people use these daily

MSFTGOOGL
🎨

Creative tools

Image gen, video gen, writing tools — replacing some creative work

GOOGLADBE
📊

Data platforms

Snowflake, Databricks — serve data to AI apps

SNOWDDOG
🛡️

AI security

CrowdStrike, Palo Alto — AI protects against AI attacks

CRWDPANW

📈 Enterprise AI software stocks:

PLTR
Palantir Technologies
AI platform for governments and big companies. Huge growth.
CRM
Salesforce
Einstein AI + Agentforce. Biggest CRM + biggest AI distribution.
NOW
ServiceNow
AI in IT, HR, customer service workflows. Strong platform play.
SNOW
Snowflake
Data cloud — critical for AI apps to access enterprise data.
DDOG
Datadog
Observability for AI infrastructure. Monitors what AI is doing.
PATH
UiPath
Robotic process automation + AI. Automates business tasks.
APP
AppLovin
AXON AI drives mobile ads. Quarterly revenue +66%. Huge margins.
TEM
Tempus AI
AI platform for medicine/oncology. 83% revenue growth.
CRWD
CrowdStrike
AI-native cybersecurity. Charlotte AI assistant.
PANW
Palo Alto Networks
AI-powered network security. Platformization strategy.
🏆

AI Winners in Software

  • Companies with distribution (Microsoft, Salesforce)
  • Companies with data (Snowflake, Palantir)
  • AI-native startups (Tempus, AppLovin)
  • Infrastructure for AI apps
📉

AI Losers in Software

  • Commodity SaaS that AI can replicate
  • IT services firms (ACN, INFY) — AI automates coding
  • Freelance platforms (UPWK, FVRR)
  • Traditional consulting firms
💡Key Takeaway
Enterprise AI is where models turn into money. The winners have distribution (Microsoft to 400M+ users) or proprietary data (Palantir, Tempus). The losers are companies whose business is selling human labor that AI can now do — like IT outsourcing (ACN, INFY) and freelance coding platforms.
🤖
Chapter 9

The Future — AI Agents & Robots

Today AI answers questions. Tomorrow it'll DO tasks. That's the biggest shift yet.

Right now, when you use ChatGPT, you have to tell it what to do each step. AI Agents change that. They can plan, take actions, use tools, and finish whole tasks on their own. In 2026, 67% of Fortune 500 companies already have at least one AI agent running in production.

🦾

What's an AI agent?

Imagine an AI that can research a topic on the internet, write a report, email it to your team, update your calendar, and book you a flight — all from you saying "handle my trip to Tokyo." That's an agent. It doesn't just CHAT — it DOES.
Infographic

🤖 How an AI agent thinks and acts

👤User gives task🧠AGENT💾Memory🔧Tools🌐Web📧APIs🛡️SafetyTask done!

🤖 The Agent Stack

🧠
Reasoning
The brain — plans what to do
💾
Memory
Remembers past info and conversations
🔧
Tools
Uses web, code, emails, databases
🛡️
Safety
Makes sure the agent behaves

Agents need MORE compute than chatbots

Here's the thing investors love: a regular chat uses AI once. An agent that researches, reasons, uses tools, and checks its work might use AI 10-100 times per single task. That's why everyone is buying GPUs like crazy. Agent adoption is a massive accelerator for GPU demand.

🦾 Robots are coming too

Physical robots powered by AI — humanoid robots, warehouse robots, self-driving cars — are next. Figure AI (private) is valued at $39 billion. Tesla is building Optimus. By 2027-2028, expect to see robots deployed in factories and warehouses at scale.

📈 Robotics stocks:

TSLA
Tesla
Optimus humanoid robot + FSD self-driving. Huge real-world AI data.
ISRG
Intuitive Surgical
Da Vinci surgical robots. AI-enhanced precision surgery.
ROK
Rockwell Automation
Industrial automation + AI for factories.
ABB
ABB Ltd
Industrial robots + electrification. Both AI tailwinds.
💡Key Takeaway
Agents are a massive multiplier on AI compute demand (10-100x per task). The shift from "AI that answers" to "AI that does" could be the most valuable transition in tech. WatchPLTR, NOW,CRM for enterprise agents, andTSLA for the physical AI story.
💰
Chapter 10

Follow the Money — Who Wins?

Let's put it all together. Here's how money actually flows through the AI chain.

Every time someone uses AI, money flows through the chain we just learned about. Here's the actual path a dollar takes:

Infographic

💸 Follow the dollar: river of AI money

$$$$👤You💻App☁️Cloud🔲GPUs🏭FactoryPowerEvery $1 spent on AI flows through this whole chain

💸 The money trail

👤
User
Asks AI
💻
App
MSFT, CRM
☁️
Cloud
MSFT, AMZN, CRWV
🔲
GPU
NVDA, AMD
🏭
Factory
TSM, ASML
Power
CEG, VST
🏆

If AI keeps growing, who wins the most?

The answer depends on what bottleneck you believe in. Here are the 5 main bets:

  1. Bet #1: GPUs stay kingNVDA wins, then TSM, ASML, SK Hynix
  2. Bet #2: Power is the bottleneckCEG, VST, GEV, ETN win
  3. Bet #3: AI infrastructure picks and shovelsVRT, ANET, APH
  4. Bet #4: Distribution winsMSFT, GOOGL, META
  5. Bet #5: Enterprise AI explodesPLTR, CRM, NOW, SNOW
🛡️

Safer bets (clearer moats)

  • NVDA — CUDA ecosystem lock-in
  • TSM — leading-edge monopoly
  • ASML — EUV monopoly
  • CEG — actual nuclear plants
  • MSFT — Office distribution
🎲

Higher risk/reward

  • CBRS — new IPO, unproven
  • OKLO — years until revenue
  • TEM — small but growing fast
  • Small AI software names
  • Anything pre-IPO
💡Key Takeaway
The boring infrastructure plays have been the best AI investments so far. Everyone knowsNVDA and it's crushed it. But the "picks and shovels" —TSM, ASML,VRT, CEG — offer less hype and stronger moats. Not investment advice — do your own homework!
🌍
Chapter 11

The Big Picture

You now understand the whole AI chain. Here's what to remember.

The 7 Big Ideas

1

AI is one big interconnected chain

From sand in a foundry to an answer on your screen — dozens of companies at each step. Break any one link and the whole chain breaks.

2

Inference will be way bigger than training

Training happens once. Inference happens billions of times. The inference market will be 10-100x bigger over time.

3

Physical bottlenecks matter more than digital ones

Power, transformers, cooling, packaging — these take YEARS to build. Code can be written in hours. The real bottleneck is atoms, not bits.

4

The "picks and shovels" are often the best bets

TSMC, ASML, SK Hynix, Vertiv, Constellation — these boring infrastructure companies have real moats and huge backlogs.

5

Nuclear went from forgotten to hot

AI data centers need 24/7 reliable clean power. Only nuclear provides that at scale. Meta signed 6.6 GW of nuclear deals in 2025.

6

Agents are the next big wave

Today AI chats. Tomorrow AI works. Agents use 10-100x more compute per task. This is accelerating the whole hardware chain.

7

Disruption creates winners AND losers

Every business AI helps, some businesses AI hurts. IT services (ACN, INFY), freelance platforms (UPWK), commodity SaaS all face pressure.

🎉 You did it!

You now understand the complete AI chain — from silicon to software, chips to agents, power to platforms. When you hear news about AI, you'll actually know why it matters and who benefits.

⚠️ Educational purposes only. Nothing on this page is investment advice. Stocks mentioned are illustrative examples of companies in the AI ecosystem. Always do your own research and consult a financial advisor before investing.