The AI
Stack Map

The internet did not simply produce websites. It produced a stack. Physical networks, routing, protocols, databases, servers, operating systems, browsers, search, payments and applications each had their own economics. The people who understood where one layer ended and another began could see why Intel, Cisco, Broadcom, Oracle, Microsoft, Google, Amazon and Apple were not all the same kind of company. They were claims on different choke points.

AI is now at the same stage. The public conversation is still trapped at the application layer: which chatbot is better, which assistant writes cleaner code, which image generator looks more realistic. That is the wrong level of abstraction. The more important question is not what the model can do today. It is which layer of the system compounds as the models get cheaper.

The first useful frame is that AI has become a six-layer stack. From the bottom up, the layers are infrastructure, chips, data, models, execution and applications. The stack then forks. One side is software AI, where inference prices are falling fast and intelligence begins to behave like abundant compute. The other side is physical AI, where the bottlenecks are not tokens but batteries, motors, sensors, materials, factories and energy.

The first layer of AI is infrastructure. That sounds boring until the constraint binds. Then it becomes the whole market.

The IEA says capex from five large technology companies rose to more than $400 billion in 2025 and is set to rise another 75% in 2026. That number is not just a symbol of corporate ambition. It is the bill for turning intelligence into an industrial utility.

Data centres require land, substations, grid interconnection, power-purchase agreements, backup capacity, cooling systems, transformers, water or liquid-cooling loops, fibre and permitting. The model layer can iterate every week. The power layer cannot.

The model layer is the loudest layer because it produces the magic. It is also the layer where rent may compress fastest.

Epoch AI's analysis of inference-price trends shows why. It found that the lowest API price required to reach GPT-4-level performance on a PhD-level science benchmark fell around 40× per year, while decline rates across benchmark thresholds ranged from 9× to 900× per year. The exact rate depends on task, threshold and benchmark. The strategic implication is clear: capability-adjusted intelligence is getting cheaper very quickly.

That creates a paradox. The model layer is becoming more important to the world but less obviously defensible as a standalone profit pool. If every frontier model becomes better, and if open-weight or cheaper closed models approach the same practical performance on many tasks, then the value above the model shifts to distribution, context and execution. The model becomes the engine. The business is the machine that uses it.

Most AI maps jump from models to applications. That misses the most important emerging layer: execution. It is what sits between a model that can reason and an application that can do work — agents, tool use, workflow orchestration, memory, permissions, identity, sandboxing, evaluation, observability, compliance, exception handling and handoff to humans.

The Federal Reserve's 2026 note on AI adoption shows that diffusion has already moved beyond novelty. By year-end 2025, about 18% of US firms had adopted AI according to Census business survey data; work-related generative-AI adoption among individuals was about 41% in November 2025; and a Survey of Business Uncertainty estimate found that roughly 78% of the labour force worked at firms that had adopted AI. But adoption is not the same as transformation. A company can have employees using AI and still have no execution layer.

The most important structural split appears at the chip layer. Above chips, AI forks into two worlds.

The first is software AI. Agents, copilots, model routing, code generation, enterprise automation, search, media, customer support, analytics and personal assistants. It compounds on declining inference prices.

The second is physical AI. Robots, autonomous vehicles, drones, industrial automation, humanoids, warehouse systems, embodied assistants. It compounds more slowly because atoms do not scale like tokens. Physical AI has to solve batteries, motors, actuators, sensors, safety, maintenance, manufacturing yield, repair networks, regulatory exposure and real-world edge cases.

AI in 2026 should be understood as a stack whose value is migrating away from novelty and toward bottleneck control. The bottom of the stack is becoming more important because intelligence consumes physical resources. Power, cooling, grid access, chips, lithography, high-purity materials and critical minerals are not background inputs. They are the reason some companies can scale and others cannot.

The chip layer turns the AI boom into a geopolitical supply-chain problem. The model layer will keep improving, but capability-adjusted inference prices are falling fast enough to pressure standalone model margins. The data layer becomes more valuable when it is private, contextual and tied to outcomes. The execution layer becomes the bridge between reasoning and work. The application layer wins when distribution becomes data and data becomes execution.

The fork matters because software AI and physical AI do not compound the same way. Software AI rides the collapse in inference cost. Physical AI rides manufacturing scale and supply-chain control. One waits on permissions. The other waits on batteries, motors, factories and minerals.

The honest version is this. AI is not one race. It is a layered contest over scarce control points. The market will remember the famous model names, but the durable winners may be the companies that own the less glamorous boundaries: powered land, lithography, high-purity quartz, memory bandwidth, proprietary workflow data, agent execution, identity, evaluation and physical supply chains.