Who Captures AI Economics
Value created and value captured are different quantities at every layer of the AI stack — and the layer that keeps the economics is decided by where substitution fails, not by position in the stack.
The most repeated forecast about AI economics is an analogy: in the internet era, the infrastructure builders overbuilt and commoditized while the value migrated up the stack to applications and aggregators — therefore the same rotation is coming for AI, and today's infrastructure margins are tomorrow's application margins. The analogy is not wrong about the internet. It is wrong about what made the internet turn out that way. Value did not move up the stack because stacks have a natural direction of flow; it moved because the layer below commoditized, and whether that happens again is precisely the open question, not the settled premise.
Here is the framing we use: value pools where substitution fails. At every interface in the stack — silicon to compute, compute to model, model to application, application to end user — there is a buyer and a seller, and the economics split according to what the buyer can do if the seller raises price. Where the buyer has alternatives, margin compresses to the cost of capital regardless of how essential the layer is; where the buyer has none, margin accumulates regardless of how thin the layer looks. Essentiality is not the test — bandwidth was essential and earned nothing. Scarcity of acceptable substitutes is the test, and it has to be re-examined interface by interface, era by era, because it is a property of market structure, not of technology.
Walking the stack, interface by interface
Apply the substitution test at each layer and the current shape of the profit pool stops being mysterious. At the silicon layer, the buyer of frontier training compute has faced, for most of the cycle, effectively one qualified seller once software ecosystem, networking, and delivery timelines are counted — so that is where the pool has sat. The important question is not whether alternatives exist on a specification sheet but whether they are substitutes at the moment of purchase: an alternative that requires re-architecting a software stack and re-qualifying a supply chain is not a substitute this budget cycle; it is a threat next cycle. That distinction — substitute now versus substitute eventually — is what separates current economics from terminal economics, and confusing the two is how investors overpay at monopoly multiples for margins that are structurally in the process of being competed toward oligopoly levels.
One layer up, the cloud and compute providers sit in an intermediate position: differentiated against small buyers who cannot build, commoditized against the largest model developers who can credibly build or multi-source, and increasingly disciplined by a fringe of specialized capacity providers. The model layer is where the substitution test is most unstable. When one model is clearly most capable, its developer holds pricing power over every application that needs frontier performance. When several models converge into rough parity — and the history of software suggests capability leads compress as techniques diffuse — models begin to commoditize each other, and the pool drains out of the model layer in both directions: down to whoever supplies the scarce compute beneath, and up to whoever owns the customer relationship above. A model in the middle of parity is the classic squeezed layer: expensive to build, cheap to substitute.
At the application layer, the test inverts. The application does not need to be irreplaceable in absolute terms; it needs to be irreplaceable to its customer, which is a claim about workflow depth, proprietary data and context, integration surface, and the organizational cost of switching — the same moats software has always had. What is genuinely new in this era is a predator at the interface below: an application whose entire value is a thin adaptation of a general model's capability can be absorbed by the model layer as capabilities broaden. The durable application moat is the part of the product the model cannot subsume by getting smarter — the distribution, the data rights, the workflow lock, the compliance surface. An application whose moat is a clever prompt is a feature on lease.
The layer nobody owns: value captured by users
The substitution framework has a corollary that rarely makes it into equity narratives: some technologies create enormous value and allow almost none of it to be captured by vendors, because competition at every layer passes the gains through to the buyer. Air travel and electricity both transformed the economies that adopted them while delivering unremarkable cumulative returns to most of the capital that built them — the value was real and it landed with users, as lower prices and expanded possibility, rather than with owners, as margin. There is a coherent scenario in which much of AI's value takes this path: capability becomes abundant, layers discipline each other, and the surplus shows up as productivity in every industry except the AI industry. For a portfolio owner this is not a curiosity; it is the scenario in which the best AI exposure is not the stack at all, but the businesses whose cost structures the stack deflates. A complete view of AI value capture prices all three destinations: captured by a stack layer, competed away between layers, or passed through to users entirely.
The illustration: two eras, two directions of migration
The internet era and the mobile era are the cleanest paired history lesson, because the same investors watched value migrate in opposite directions within a decade. In the internet buildout, the physical layer commoditized catastrophically: bandwidth was durable, overbuilt, and interchangeable, so its price collapsed toward marginal cost and stayed there. With the layer below commoditized, the economics pooled at the layers that aggregated demand on top of cheap infrastructure — search, marketplaces, social platforms. Investors who generalized this into a rule — infrastructure commoditizes, applications win — then watched the mobile era do the reverse. In mobile, the integrated platform layer — hardware fused with operating system fused with distribution — never commoditized. It taxed the application layer through its store, held the customer relationship, and kept the deepest profit pool of the era at what internet-era logic would have called the infrastructure level.
The difference was not fashion; it was where substitution failed. Bandwidth had perfect substitutes; the integrated mobile platform had none. AI will obey the same rule, whichever way it resolves. If frontier compute behaves like bandwidth — durable, overbuilt, interchangeable — the pool migrates upward as the analogy crowd expects. But frontier compute differs from fiber in exactly the properties that made fiber commoditize: it depreciates on a short competitive clock, requires continuous reinvestment at the frontier, and carries an ecosystem attached to it. A glut of obsolete compute does not discipline the price of frontier compute the way a glut of immortal fiber disciplined bandwidth. That does not guarantee the pool stays down the stack — it means the question is genuinely open, which is the one position the confident analogies on both sides refuse to hold.
The strongest case against this framework
The most serious objection comes from the aggregation school: distribution, not substitution, decides value capture. On this view, whoever owns the user relationship at scale eventually commands the stack, because every layer below competes for access to demand that only the aggregator controls — and the history of consumer technology largely vindicates this. The objection continues: substitution analysis is static, always describing the previous bottleneck, while distribution advantages compound quietly until they set the terms for everyone.
We think this is less a rival framework than a special case of the same one. Owning distribution is a substitution failure — at the customer interface, the most valuable interface in the stack. The aggregation view is a bet that the customer-relationship moat decays slower than the compute and capability moats below it; often true, but not free, and AI attacks it in a novel way: an assistant that mediates tasks for the user is a bid to become the new point of aggregation, which makes the current distribution moats contestable in a way they have not been for a long time. Where we do concede ground is on dynamism — a substitution map is a snapshot, and it must be re-drawn as leads converge and ecosystems get bridged. That is how Sterling's Embedded Intelligence treats the problem: not as a one-time verdict on which layer wins, but as a standing set of interface tests — compute commoditization, model convergence, application switching costs, aggregation contests — each of which is observable, and each of which moves the pool when it flips.
How to apply this framework
- Run the substitution test at the interface, not the layer. For each holding, name its buyer and ask what that buyer does on a price increase — and whether alternatives are substitutes this budget cycle or only eventually. Current margin prices the first; your terminal assumption must price the second.
- Watch model convergence as the pool's main valve. Durable capability leads keep economics at the model layer; parity drains it downward to scarce compute and upward to owned distribution. Evidence of convergence or divergence at the frontier moves every layer's terminal margin at once.
- Grade applications by what a smarter model cannot subsume. Workflow depth, proprietary data rights, integration and compliance surface, owned distribution — these survive capability growth beneath them. A moat that is only a head start on using the model is a lease, and should be valued like one.
- Price the pass-through scenario. Some or much of AI's value may be captured by no vendor — competed through to users as surplus. Exposure to the businesses whose costs the stack deflates is a legitimate AI position, and in the pass-through scenario it is the best one.