Overview
AI is generating staggering revenue, but revenue isn't profit, and profit isn't evenly captured. To understand who actually makes money, follow the value chain layer by layer. The old gold-rush wisdom applies: often the shovel-sellers win.
The layers of the AI stack
Chips: The clearest winners so far are the makers of AI accelerators and the supply chain around them. Demand vastly exceeds supply, so they command pricing power and fat margins — the literal shovel-sellers.
Cloud/infrastructure: The hyperscalers renting GPUs and AI services capture durable, recurring revenue. They own scarce capacity and existing customer relationships.
Model labs: Frontier labs earn enormous revenue but also spend enormous sums on training and inference. Margins are pressured by competition and compute costs, and capability leads are short-lived. Big top line, uncertain bottom line.
Applications: Thousands of app companies sit on top. A few with real moats thrive; many are thin "wrappers" with weak differentiation, thin margins, and exposure to being copied or absorbed.
Where durable value accrues
Profit concentrates where there's scarcity and defensibility: scarce chips, scarce capacity, proprietary data that compounds, distribution advantages, and deep workflow lock-in. Layers without scarcity — undifferentiated apps reselling the same model — get competed to thin margins.
Don't forget consumer surplus
A huge share of AI's value isn't captured as profit at all — it flows to users as consumer surplus (faster work, cheaper services). That's great for society but means "AI is valuable" doesn't translate cleanly into "everyone in AI is profitable."
What this means for you
If you invest or build: ask where the scarcity and moat are in your layer. Infrastructure and data/distribution-rich applications capture value; commodity layers don't. If you use AI: recognize that much of the benefit accrues to you, the user — use it aggressively.
Honest limits
The chip/infrastructure dominance reflects today's build-out phase; value can shift down the stack as the market matures (as it did in past tech cycles). The durable lens isn't "which layer" but "where is the scarcity" — and that can move.
