Follow the Money — The AI Value Chain from Chip to Chatbot

1/ An AI chatbot answering your question looks simple. Behind it is a supply chain spanning continents, worth trillions, with value captured at every layer. Follow the money from chip to chatbot. 🧵

2/ LAYER 1 — MINERALS: It starts in mines. Lithium (Chile, Australia), cobalt (DRC, where child labor persists), rare earths (China controls ~60% of mining, ~90% of processing). The physical foundation of AI is extractive industry.

3/ LAYER 2 — CHIP DESIGN: Nvidia dominates AI chip design with 80-90% market share. In 2024, Nvidia's revenue hit $130B with net margins around 55%. The single most profitable chokepoint in the AI stack. Jensen Huang's leather jacket prints money.

4/ LAYER 3 — FABRICATION: Nearly all advanced AI chips are made by TSMC in Taiwan. One company, one island, one earthquake zone. TSMC's margins are ~40%. The world's most consequential factory you've never visited.

5/ LAYER 4 — CLOUD INFRASTRUCTURE: AWS, Azure, and Google Cloud rent GPU clusters for $2-4 per GPU-hour. A frontier model training run can cost $500M+ in compute alone. Cloud margins run 25-35%. The landlords of AI.

6/ LAYER 5 — MODEL TRAINING: This is where the labs live. OpenAI, Google DeepMind, Anthropic, Meta spend billions training frontier models. Most are not yet profitable. They capture value through licensing, APIs, and the promise of future dominance.

7/ LAYER 6 — DATA LABOR: Thousands of workers — mostly in Kenya, India, the Philippines — label data and rate outputs for $1-3/hour. This labor makes AI "work." It captures almost none of the value it creates. The base of the pyramid.

8/ LAYER 7 — APPLICATION: Companies like Jasper, Harvey, Cursor build products on top of foundation models. They capture value through vertical expertise, but they're dependent on the layers below. If OpenAI raises API prices, their margins evaporate.

9/ LAYER 8 — DISTRIBUTION: Microsoft bundles Copilot into Office. Google embeds Gemini in Search. Apple builds AI into iOS. Distribution is the ultimate moat. The companies with existing user bases capture disproportionate application-layer value.

10/ THE MARGIN MAP: Nvidia ~55%. TSMC ~40%. Cloud providers ~30%. Model labs: mostly negative (burning capital). Data workers: subsistence. Application builders: 10-20%, fragile. The value flows to hardware and infrastructure. Not to labor. Not to users.

11/ Notice the pattern: the closer you are to the physical chokepoints (chips, fabs, clouds), the more value you capture. The closer you are to human labor (annotation, moderation, support), the less. AI reproduces capitalism's oldest distribution rule.

12/ Who pays, who profits? Miners, data workers, and ratepayers bear costs. Nvidia, TSMC, and cloud giants capture value. Users get productivity. The redistribution question: is this arrangement necessary, or just familiar? /end

---

LinkedIn version:

An AI chatbot answering your question looks simple. Behind it is a supply chain spanning continents, worth trillions, with value captured unevenly at every layer.

The chain starts in mines — lithium from Chile, cobalt from the DRC, rare earths from China. Nvidia designs the AI chips (80-90% market share, ~55% net margins on $130B revenue). TSMC in Taiwan fabricates nearly all of them (~40% margins). Cloud providers — AWS, Azure, Google Cloud — rent GPU clusters at 25-35% margins. They're the landlords of AI.

Then come the model labs: OpenAI, DeepMind, Anthropic, Meta spending billions on training, most not yet profitable. The data workers who make these models function — labeling data and rating outputs in Kenya, India, the Philippines — earn $1-3/hour and capture almost none of the value they create.

At the application layer, companies build products on foundation models with fragile 10-20% margins. And at the distribution layer, Microsoft, Google, and Apple bundle AI into existing products, leveraging their user bases as the ultimate moat.

The pattern is unmistakable: the closer you are to physical chokepoints (chips, fabs, clouds), the more value you capture. The closer you are to human labor, the less. AI reproduces capitalism's oldest distribution rule.

The redistribution question isn't just who benefits from AI — it's whether this value distribution is necessary or merely familiar.