Who Captures the Value in the AI Stack?
4 min read · updated August 3, 2026
Value created and value captured are different quantities, and a technology can create an enormous amount of the first while almost none of the second reaches the people who built it. Which layer keeps the difference is decided by substitutability, and substitutability can be interrogated layer by layer.
What capturing value means
A layer captures value when it can charge more than its cost without losing the business. That requires two things at once: its buyers have few alternatives, and it has alternatives for its own suppliers. A layer squeezed on both sides — many equivalent competitors, concentrated suppliers — passes value through in both directions and keeps very little, however essential it is.
Being essential is not the criterion, which is the mistake this analysis exists to prevent. Electricity is essential to every business in the economy and utilities do not capture the value of the economy. Necessity without scarcity earns cost plus a normal return.
It helps to hold the three quantities apart explicitly. Value created is what the buyer gains — the work done, the time saved. Value captured is the part that ends up as somebody’s profit. The remainder goes to buyers as consumer surplus, and it is usually the largest of the three in a competitive market. A technology can therefore be enormously important and financially disappointing for almost everyone who builds it, and this is a normal outcome rather than a paradox: competition is the mechanism that transfers created value to buyers, and it works.
This is also why the layer analysis has to be done per layer rather than for the industry. Asking whether AI is profitable is asking about eight different businesses with different supplier structures, and the answer can be yes at one layer and no at the one immediately above it, simultaneously, for the same underlying activity.
The four questions
- How many viable suppliers does a buyer at this layer have? Not how many exist — how many the buyer could actually qualify and switch to within a planning cycle.
- What does it cost to switch between them? Including requalification, retraining, migration of accumulated state, and the risk of a silent regression.
- What share of the buyer’s total cost is this layer? A large share attracts effort to reduce it — second-sourcing, in-housing, redesign. A small share is tolerated even when the price is high, which is why small components in a big system often earn well.
- How fast is entry? The relevant barrier is time and capital together: how long from a decision to fund an entrant to that entrant taking a real order.
The four are enough because they capture both sides of every trade in the stack. A layer scoring well on all four — few substitutes, expensive to switch, small share of the buyer’s spend, slow entry — is where value pools.
The layers, question by question
What follows is the questions applied, not an assertion about what each layer currently earns. Where an answer has changed within recent memory it is flagged, because that is the point of doing it this way.
| Layer | Description |
|---|---|
| Energy and land | Suppliers are few in any given location, switching is impossible without moving the building, and lead times are measured in years. Share of total cost is modest but rising with scale. Entry is slow for physical and regulatory reasons that no amount of capital compresses. |
| Fabrication and advanced packaging | Very few qualified suppliers, requalification is expensive and slow, and entry requires capital plus process knowledge that takes years to accumulate. On the four questions this is about as concentrated as a layer can be. |
| Accelerators and their software | Substitutes exist on paper; the switching cost is in the software ecosystem rather than the silicon, which is the durable part. Share of buyer cost is large, which is exactly why buyers invest so heavily in second-sourcing and in their own designs. |
| Cloud and datacentre operation | Several credible suppliers, real but manageable switching costs, and a large share of buyer cost — which means constant pressure. The distinguishing asset is having capacity when others do not, and that is a capital and lead-time position rather than a technical one. |
| Model training | The number of suppliers clearing any fixed capability bar rises over time, and open weights push the floor price of that bar toward zero. Switching cost is low at the API and high in evaluation and prompt-fitting. Entry is expensive at the frontier and cheap below it. |
| Orchestration, gateways and tooling | Many suppliers, deliberately low switching costs — being easy to leave is often the pitch — and a small share of the buyer's spend. Value here comes from being the place where state, evaluation and routing policy accumulate, not from the routing itself. |
| Applications and workflow | Substitutes are numerous in the abstract and few in practice once records, permissions and integrations exist inside a customer. Switching cost is the highest anywhere in the stack, and entry is fast — which is why this layer is simultaneously crowded and sticky. |
| Distribution and the end relationship | The scarcest thing in the diagram: attention and an existing installed base. Not becoming cheaper, not purchasable in bulk, and it accumulates through ordinary operation. |
Why the answer migrates
The answers above are not stable, and the direction of drift is predictable. Any layer whose scarcity is technical gets less scarce as the technique diffuses. Any layer whose scarcity is physical — land, power, fabrication capacity, a building with a grid connection — stays scarce as long as the lead time stays long.
The general rule is that value migrates toward whatever cannot be scaled by writing software, and it migrates away from anything that can. That is uncomfortable for the layers that feel most important, and it is consistent with what has happened in older stacks: the components people found most intellectually interesting rarely captured the most.
The second regularity worth carrying: when a layer commoditises, the value released does not evaporate. It moves to an adjacent layer, usually the one that was previously bottlenecked by the cost of the now cheap one. Ask, whenever a layer gets cheap, which layer is now the constraint — that is where the value went.
There is a third force worth watching, which is that layers do not stay separate. Firms integrate across them: a model developer builds applications, an application company trains its own models, a cloud designs its own accelerators. Each such move is an attempt to escape a squeeze, and integration succeeds when the acquired layer’s scarcity was the thing constraining the firm and fails when the layer was already competitive — in which case the firm has bought itself a business it now has to run at market returns while its rivals simply keep buying the input. Watching who integrates in which direction is therefore a reading of where the squeeze is, and it is a more reliable signal than any statement of strategy, because it is expensive.
Finding the binding constraint
Which layer is binding today changes, and it is knowable without insider information. Three observable signals, each of which you can check yourself:
- Lead time. The layer people wait longest for is the constraint. Waiting is the market signalling scarcity before prices fully adjust.
- Allocation instead of pricing. When a supplier rations rather than raising the price — queues, approvals, reserved capacity for favoured buyers — that is a strong indicator of a genuine shortage.
- Where the vertical integration is aimed. Firms integrate toward whatever is squeezing them. Watch which layer large buyers are trying to build or buy their way into; that is their answer to this question, backed with capital.
This page deliberately names no company and gives no figures for what any layer earns. Those numbers exist in filings with dates on them, and the questions above are what turns such a number into an explanation rather than a fact you have to keep re-checking.