The AMICE Stack and the Far Side of Model Mountain
Applications, Models, Infrastructure, Chips, Energy. Most of us will spend our whole AI lives on the top layer. Here's what's holding it up.
2026-08-12
Loading content...
Applications, Models, Infrastructure, Chips, Energy. Most of us will spend our whole AI lives on the top layer. Here's what's holding it up.
When people argue about AI they're arguing about apps and models, but the money and the geopolitics live in a five-layer stack: Applications, Models, Infrastructure, Chips, Energy. Most of us will spend our entire AI lives on the top layer, and that's fine, as long as we know the mountain is there.
When most people talk about AI, they're talking about two things: the app they typed into this morning and the model behind it. ChatGPT versus Claude. Which one writes better emails. That's the whole conversation.
But when economists and governments talk about AI, they're talking about something much taller. The shorthand I like is the AMICE stack: Applications, Models, Infrastructure, Chips, Energy. Five layers, top to bottom, and every one of them is a real industry with real winners and losers. Jensen Huang has been drawing this same picture in his keynotes: energy at the bottom, chips above it, then the data centers, then the models, then the apps where AI finally meets a person. His point is that every prompt you send pulls on every layer beneath it, all the way down to the power plant.
Applications are where you live. Chat apps, coding agents, the AI feature inside your email client. This is the layer with users in it.
Models are the intelligence itself. GPT, Claude, Gemini, the open-weight families. Training one at the frontier costs hundreds of millions of dollars, which is why the list of companies doing it is short.
Infrastructure is the data centers. Land, buildings, cooling, networking, and the orchestration that makes a hundred thousand GPUs act like one machine. This is where the hyperscalers spend their capex, and the numbers have gotten so large they show up in GDP conversations.
Chips are the silicon underneath. NVIDIA's GPUs, Google's TPUs, and the one company that fabricates nearly all of them: TSMC. This layer is why Taiwan sits at the center of AI geopolitics, and why Dario Amodei spent an entire essay arguing about export controls. Chips are the layer nations fight over.
Energy is the floor. Training runs and inference farms consume power at a scale that has utilities and regulators scrambling. The IEA has a whole report on energy and AI because data center demand is now a grid-planning problem, not a footnote.
The layers interact in one direction going down and one going up. Demand flows down: a popular app needs more inference, which needs more racks, which need more chips, which need more megawatts. Constraint flows up: if the grid can't deliver power, the data center doesn't get built, the chips sit in a warehouse, the model doesn't train, and your app gets rate-limited. Sam Altman takes this to its logical end when he says compute is going to be the currency of the future. Satya Nadella went further at Davos and named the metric: tokens per watt per dollar. How much intelligence can you produce per unit of energy and money. That's a national competitiveness number now, like steel output used to be.
Here's the part I think most people get wrong about themselves: even when you open the model picker and choose between models, you are still in the Application layer. Choosing a model from a dropdown is an application feature. You're a customer of the stack, not a participant in it. I've written before that the best AI model is the one you actually use, and I stand by it, but let's be honest about what that choice is: it's picking a dish off a menu, not cooking.
And that's fine. The Application layer is enormous, and realistically it's where most of us will spend most of our time. I picture the Model layer as a mountain ridge. On our side of Model Mountain: every app, every workflow, every prompt, billions of users. On the far side: the trenches. Chip fabs, interconnects, transformer substations, training runs. A lot of the fastest-moving AI work is happening over there, and few of us will ever hike over the ridge to see it. You can have a full and productive AI life without ever crossing. But you should know the far side exists, because that's where the constraints that shape your side get decided.
A couple of years ago, VC circles had a favorite insult: "that's just a ChatGPT wrapper." And the insult was earned. In 2022 and 2023, a lot of funded startups were a thin UI over the OpenAI API with slightly better prompts than the next product. TechCrunch was warning founders that integration alone wasn't a business. Millions got raised anyway. Most of it turned no real profit, and when OpenAI shipped a feature or changed a price, whole companies evaporated.
The lesson people took from that era was "don't build at the Application layer." That's the wrong lesson. The right lesson is that the Application layer rewards the same things software always rewarded: distribution, workflow depth, and owning a customer problem end to end. A wrapper fails because it's shallow, not because it's an application. The slur itself has aged badly enough that Forbes ran a piece arguing only fools use "ChatGPT wrapper" as an insult now. Cursor is a wrapper by the 2023 definition. So is most of the AI software making actual revenue today. The difference between a wrapper and a product turned out to be everything around the API call.
Once you have the five layers in your head, AI news stops being a blur. Every headline is about one layer, and the layer tells you who it affects.
A new model release is an M story, and unless you build apps, it reaches you only after an A-layer product adopts it. An export control announcement is a C story that will show up as an M constraint in about two years. A utility signing a nuclear deal with a hyperscaler is an E story that decides which I-layer buildouts happen at all. When someone predicts AI progress will stall, ask which layer they think stalls. Usually they haven't picked one, which means it isn't a prediction, it's a mood.
Looking forward, I'd frame it this way. The top of the stack keeps getting more crowded and more competitive, because it's the cheapest layer to enter. The bottom three layers keep getting more concentrated and more political, because they're the most expensive layers to enter, and the countries that control them know it. The interesting careers and the interesting investments over the next decade are at the boundaries: people who can stand at the Application layer and reason about energy contracts, or stand at the Infrastructure layer and understand what app developers actually need.
Most of us will keep living happily on our side of Model Mountain. Just remember that when your app feels slow, or your subscription price goes up, or your favorite model gets rate-limited, the reason usually isn't on your side of the ridge. It's three layers down, in a trench somewhere, measured in watts.