Where value accrues in a stack
A product reaches the customer through a chain of layers — components, assembly, software, distribution. The total profit gets divided among those layers, but never evenly. The rent collects where supply is scarce and at the layer that owns the customer; the substitutable middle commoditizes. The question for any technology shift is simply: which layer keeps the money?
The smiling curve
Plot value captured against position in the chain and you often get a smile: high at the two ends, low in the middle. Upstream — scarce inputs, IP, hard-to-build components — captures value because it’s hard to substitute. Downstream — brand, distribution, the customer relationship — captures value because it owns demand. The middle, where the work is real but replaceable (assembly, undifferentiated processing), gets squeezed.
In PCs, the ends — Intel (CPU) and Microsoft (OS), plus the retail relationship — kept the profit; box assembly was a brutal low-margin commodity. The hardware in the middle did the most visible work and earned the least.
Why: scarcity and the customer relationship
Two forces decide where rent lands:
- Scarcity. A layer with few credible suppliers and high barriers (capital, IP, process) can hold price. A layer anyone can provide can’t.
- Ownership of demand. The layer the customer actually chooses and pays — the brand, the interface, the relationship — can dictate terms to the layers behind it.
Helmer’s 7 Powers is the rigorous version of “scarce”: scale economies, network economies, switching costs, counter-positioning, cornered resource, branding, process power. A layer keeps its margin only if it has at least one. A layer with none is the commodity middle — no matter how essential its function.
Applied to the AI stack
Lay the AI stack end to end and ask where the smile sits:
- Upstream (scarce): leading-edge fabs, HBM, the GPU platform + its CUDA lock-in, power. Few suppliers, brutal barriers — strong value capture.
- The model layer (contested middle): real scale economies, but weak network effects, low switching costs, and an open-weight price floor. This is the layer most at risk of being the commodity middle — and the layer rivals are actively trying to commoditize.
- Downstream (owns demand): the products and distribution surfaces — ChatGPT, Copilot, agents, the device and the channel. Whoever owns the customer relationship and proprietary context can hold margin.
Hence the through-line of the model-layer series: the rent is migrating to the ends — scarce compute upstream, owned distribution downstream — and a pure model vendor sitting in the middle is in the squeezed position. The winning labs escape the middle by becoming a downstream product (owning the customer) or by holding a scarce upstream edge.
How to use it
- Draw the whole chain from raw input to end customer.
- Mark each layer scarce or substitutable — does it have at least one of the 7 Powers?
- Find who owns the customer. That layer can squeeze the others.
- The squeezed middle is whoever has neither. That’s where margin goes to die.
The capstone of the moat primers — it composes network effects, scale economies, and switching costs into a single question: which layer keeps the money? Any “who captures the value” argument links here.