If you don't understand this chart, you don't understand the AI space.
When I talk AI with leadership teams, the same thing happens. They know the apps. They've heard of Nvidia. Everything in between is a blur.
So I drew it as a mine. Every layer wraps the one below it. The app you use sits on a model, which runs in a cloud, which rents data centers full of GPUs, which TSMC makes on ASML machines, from refined silicon and metals, powered by a lot of electricity. 15 layers, bottom to top. Watch the cage in the shaft. Every app rests on everything below it. The software runs on compute. Compute depends on the factories and energy beneath.
The boxes on the left show company gross margins for ASML, TSMC and Nvidia, plus rough ranges for labs and apps. They're company figures, so read them as a guide to where the money pools. The chokepoints sit at the bottom.
The foundation (layers 1 to 6)
This is the part almost nobody looks at. It's also where the whole stack can break.
Energy. Data centers will use 945 TWh by 2030. AI is the main driver. Every layer above runs on it.
Raw materials. China makes 99% of primary gallium. Chips need it.
Refining. China makes 93.5% of polysilicon, most of it solar grade. Chips need an even purer version.
Machines. Only ASML makes EUV lithography machines. One company in the Netherlands. ASML ran a 57.5% gross margin (Q3 26).
Design tools. Three firms hold 85%+ of chip design software.
Fabs. TSMC takes 72.5% of global foundry revenue. TSMC ran a 66.4% gross margin (Q3 26).
Look at how few names there are. One machine maker. One foundry with most of the market. One country for most of the gallium and the polysilicon. The supply risk for all of AI sits down here.
Compute (layers 7 to 10)
This is what AI runs on. It's also home to Nvidia, the fattest margin on the chart.
Memory + packaging. Three firms make all HBM memory, the fast memory stacked right next to the chip. SK hynix had about 59% in 2025, then Samsung and Micron.
GPUs. Nvidia made $89B from data centers in one quarter. Its gross margin was 75% (Q2 FY27). Google TPUs and AWS Trainium are the other chips in the race.
Data centers. Big Tech guided up to $725B of capex for 2026, most of it for AI. That is up 77%. Foxconn builds about 40% of AI servers.
Clouds. AWS, Azure, Google Cloud and CoreWeave rent it all back by the hour. One H100 costs $1.49 to $12 an hour depending where you rent it. CoreWeave alone sits on a $99.4B backlog.
$725B of capex is a bet that the layers above will pay for it. Every app subscription, every API call and every agent run is part of how that bet gets paid back.
This is also where the next big bottleneck sits. Demand for compute is growing faster than the world can make GPUs, find power and put up buildings. The capex is huge and a new data center takes years to build. If I had spare cash to put somewhere, it would go into layers 8 and 9. That's just my view from where I sit.
Software (layers 11 to 15)
This is what you rent. It's also where most of the noise is.
Models. Three labs took 88% of enterprise API spend in 2025. Anthropic 40%, OpenAI 27%, Google 21%. Gross margin is somewhere around 35 to 65% (an estimate).
Inference + routers. They serve many models to apps through one key. OpenRouter lists 400+ models from 60+ providers. Fireworks, Together, Groq and Bedrock sit here too.
Tools + data APIs. The data and actions agents call. There are 30,000+ MCP servers. Firecrawl, Exa, Bright Data, ZoomInfo and FullEnrich all live here.
Aggregators. One key for many tools and data providers. Clay waterfalls 100+ providers. Zapier and n8n do the same for app connections.
Apps + the economy. What people actually use. Apps took $19B of $37B enterprise GenAI spend in 2025. Cursor passed $2B a year. Lovable hit $100M ARR in about 8 months. Typical gross margin 50 to 60%.
The top layers are wrappers on wrappers. A data provider wraps other data providers. An aggregator wraps those. An app wraps the aggregator and a model. That's fine. Most software has always been built this way. The real question is what each wrapper adds.
What it means for a company buying AI
When you buy an AI tool, you buy the top of this stack. Everything under it comes with the price.
Your price comes from below. When model costs go up or GPU hours get tight, you'll see it in your renewal.
Your outages come from below. When a model provider or a cloud has a bad day, every app on top of it has one too.
Ask which layers they own. A vendor that owns its data or its workflow has something to hold on to. A vendor that is a thin layer over one model has a short shelf life.
Check where the moat is. Cursor and Lovable grew fast. Both run on models from the layers below.
Then look at the margins once more. The companies that own the chokepoints run the fattest ones. ASML 57.5%, TSMC 66.4%, Nvidia 75%. Labs sit around 35 to 65% (an estimate) and apps typically 50 to 60%. Knowing who sits where tells you who has pricing power over you.
The short version
Every layer is a wrapper on the one below. The fattest margins belong to the companies that own the chokepoints. The chokepoints sit at the bottom. Most people only look at the top. Save the chart and pull it out the next time someone pitches you an AI app.
Hope it helps. Don't hesitate to reach out on LinkedIn if you have any questions but keep the questions as specific as possible please.
David
