Follow the Dollar: Unit Economics of the AI Inference Stack, and Where Value Capture Moves Next
Every AI dollar takes a journey nobody maps. It leaves a user's credit card, passes through eleven layers, and comes to rest in surprisingly few pockets. We followed one dollar all the way down, asked who really holds the power, and found the answer every mature supply chain gives: whoever holds the customer's attention and outcomes sets everyone else's margin.
7/8/2026 · 16 min · jusCode · Read as Markdown
TL;DR
Follow one AI dollar through the eleven layers of the inference stack and it comes to rest in three pockets: the application that owns the user, the lab that owns the weights, and the chipmaker that owns the compute. The seven layers in between touch every cent and keep under two.
- Of every dollar spent on an AI product today, roughly 58 cents rests at the application, 21 at the model lab, 8 at the chipmaker, and under 2 cents across the seven layers in between.
- The iron law of supply chains: the layer that holds the end customer's attention and outcomes sets the price. Every layer above it in the chain becomes a negotiated cost line.
- That law explains the biggest strategic story of 2026: model labs storming the application layer with consumer apps, browsers, devices, and agents, because their core API product deflates tenfold a year while attention never deflates.
- Every mature digital supply chain grew tolls for its flow-governing middle: 0.15 percent for payment rails, about 3 percent for the full card stack, 15 to 30 percent for stores and marketplaces. AI's middle collects under 2 percent today. That gap is the opportunity.
- The endgame is outcome pricing: when you charge per resolved ticket or merged PR instead of per token, tenfold token deflation stops being your revenue problem and becomes your margin expansion.
- The twist ahead: when agents become the buyers, habit and brand stop working. Agents choose by price and benchmark on every call, which makes the router the new distribution.
The autopsy
One dollar, eleven layers, three pockets
A CFO sees AI as a line item. A founder sees it as gross margin. An economist sees something more interesting: a brand-new value chain settling into shape in real time, the kind of thing that happens once a decade and rewrites who gets rich. All three need the same map, and nobody has drawn it: when a dollar is spent on AI, which of the eleven layers of the inference stack actually keeps it?
So we reconstructed the journey from public reporting. The numbers below are illustrative, real value chains are messier and every company's mix differs, but the shape is the story, and the shape is unambiguous. A dollar leaves the user and comes to rest in three places: the application that owns the user, the lab that owns the frontier weights, and the chipmaker that owns the supply of compute. Everything between those three pockets is, financially speaking, a rounding error. For now.
The iron law
Power belongs to whoever holds attention and outcomes
Before touching AI, look at how every mature supply chain settles, because they all settle the same way. Grocery retailers dictate terms to the food giants whose brands fill their shelves, because the retailer owns the shopper. Apple takes up to 30 percent of app revenue from developers who did all the work, because Apple owns the home screen. Card-issuing banks take the largest slice of every swipe, because the bank owns the cardholder. In each case the pattern is identical: the layer adjacent to the end customer sets the price, and every layer upstream becomes a cost line to be negotiated downward.
What exactly does the customer-holding layer hold? Two things, and it's worth separating them. First, attention: the habit, the icon, the default, the place the customer starts. Second, outcomes: being the layer where the customer's problem is visibly solved, the merged PR, the resolved ticket, the booked flight. Attention gives you the right to charge. Outcomes give you the right to charge more than cost-plus, because you're priced against the customer's alternative (hiring a person, losing a sale), not against your own input bill. Everyone upstream of you is priced against their input bill. That asymmetry is the entire game.
Now reread the waterfall with this law in mind. The application's 58 cents isn't a reward for technical difficulty; six of the eleven layers are technically harder. It's rent on attention and outcomes. The chipmaker's 78 points of gross margin is the one legitimate exception, rent on physical scarcity, and physical scarcity is the only kind with an expiry date printed on it: fabs are being built right now. Which brings us to the most telling strategic behavior in the industry.
The tell
Why the model labs are storming the application layer
If you want to know where value capture is going, ignore what the most sophisticated players say and watch where they spend. The frontier labs, the companies with the deepest visibility into this value chain, are executing the most aggressive up-stack migration in recent business history. One turned its chat product into a consumer platform that by its own account reached hundreds of millions of weekly users, then bought a legendary hardware designer's startup for billions to build devices, then shipped its own browser. Another poured its energy into agentic coding and knowledge-work products that sit directly in the user's workflow, exactly where outcomes happen. These aren't side bets. Consumer and product revenue, not API calls, is where the growth is going.
Why would the companies that own Layer 09 fight so hard for Layer 01? Because they can read their own price sheet. Their core wholesale product, a token of fixed capability, deflates roughly tenfold per year [3], the fastest input deflation ever recorded, with open-weight models chasing the frontier from below. Selling tokens is selling a melting asset. So they're climbing to the two things that never deflate: attention (the chat app as the new home screen) and outcomes (agents that finish work, priced against labor). The migration is the industry's most sophisticated players voting, with billions of dollars, for the iron law above.
And notice the counter-migration running in the opposite direction: the strongest applications are reaching down-stack, fine-tuning and even training their own models to cut the 35-cent API line out of their waterfall. The endpoints are integrating toward each other. Which raises the obvious question: what happens to the middle they're both crossing?
The secret
Value pools where the dollar rests. It compounds where the dollar passes.
Here is the mechanism the resting map hides. Two forces are moving in opposite directions: the price of a token is collapsing tenfold a year [3], while total token consumption explodes faster, the 160-year-old pattern Jevons documented with coal, where efficiency doesn't shrink consumption, it detonates it [1]. In that economy, owning the thing that deflates is a losing position, and owning the flow is the winning one. The layers that touch every request, the gateway that meters it, the router that prices it, the cache that recycles it, hold a toll position: a small take on 100 percent of an exploding flow, at near-zero marginal cost. Payments taught this lesson already. The card networks take a fraction of a percent of everyone else's commerce and became more valuable than most of the banks whose money they move. The AI interchange layer exists today. It's layers 02 through 06, it touches every cent in the waterfall above, and the industry is currently giving it away as open source. That is the secret hiding in the unit economics.
The precedents
What every other supply chain pays its middle
If the toll thesis sounds speculative, price it against history. Every mature digital value chain grew a paid middle, and the take rates are public. Follow a hundred dollars through a card swipe: the merchant surrenders roughly two and a half to three dollars in total, of which the issuing bank, the one holding the cardholder, takes the lion's share as interchange, the processor and gateway take their slice, and the network itself, pure rails touching everything and owning no customer, takes around fifteen basis points. Tiny percentage, universal flow, legendary business. Climb the power gradient and the tolls climb with it: online travel agencies charge hotels 15 to 25 percent for owning the booking moment, food delivery platforms charge restaurants similar, app stores charge developers 15 to 30 percent for owning the install, and studies of the advertising supply chain found intermediaries absorbing on the order of a third or more of every ad dollar between advertiser and publisher.
Read the pattern, because it's the same iron law wearing different clothes: the toll you can charge is proportional to how much customer power your position holds. Pure rails with no customer relationship settle near a fraction of a percent, at planetary volume. Positions that gate demand settle at 15 to 30. AI's middle layers today collect under 2 percent while doing work, routing, governing, recycling, that in payments terms spans the gateway, the risk engine, and the network all at once. There is no historical precedent for that middle staying free. There are only precedents for how large its toll eventually gets.
The margin ladder
Who makes what: every participant in the chain, one chart
Take rates describe the middle. Gross margins describe everyone, so let's line up the whole supply chain the way a fund analyst would. Down at the physical base, the foundry that fabricates the chips runs gross margins in the mid fifties, remarkable for manufacturing, earned through a near monopoly on leading-edge process. Memory makers ride a boom-bust cycle that currently favors them, call it mid forties at the top of the cycle. The chipmaker sits above both at roughly seventy-five points, the margin of a software company charged on hardware, which is exactly what a proprietary compute runtime lets you do.
Then comes the layer everyone misreads: the GPU clouds. Their paper gross margins can look impressive, but the honest number arrives after depreciation and interest, because their product is a rapidly depreciating asset bought with borrowed money. Economically they clear something in the mid twenties, the margins of an airline wearing the costume of a software company. Model labs earn around sixty points on inference, healthy, but earned on a treadmill where the product's price falls tenfold a year. Pure API resellers scrape single digits. The middle-layer meters, our thin sliver, barely register today, though the toll table above suggests where operated meters settle once a chain matures. And at the top, the application archetypes fan out: subscription apps around sixty, seat-based enterprise tools a notch higher, and outcome-priced agents pushing past eighty, because their price is anchored to labor while their cost is anchored to a deflating token.
Read the ladder vertically and the iron law appears again in a new form: margin is not a reward for difficulty; it's a readout of position. The foundry does the hardest engineering on earth for fifty-five points. The outcome-priced app assembles rented parts for eighty. The chain pays for power over the customer and scarcity of supply, in that order, and for nothing else.
The endgame
Outcome pricing: where the power law completes
There's one more move on the board, and it's the one that ties attention, outcomes, and deflation into a single strategy. Watch what the sharpest application companies are doing to their price sheets: a leading customer-support agent charges 99 cents per resolved conversation, not per token; coding agents increasingly justify themselves in cost per merged PR; sales agents in cost per qualified meeting. This is not a pricing gimmick. It's a re-anchoring of the entire unit economics: the price is set against the customer's alternative (a support rep's salary, an engineer's hour), while the cost is set by a token bill that deflates tenfold a year.
Run that spread forward and you get the most beautiful sentence a CFO can read: when you price outcomes, token deflation stops being your revenue problem and becomes your margin expansion. Every year the input gets ten times cheaper, the outcome stays priced against labor, and the difference lands in your gross margin. This is why the labs are racing to sell agents rather than tokens, why the smartest apps are racing to define billable outcomes, and why the middle layers matter twice over: they're not just tolls, they're the machinery (routing, caching, context, verification) that turns a deflating token into a chargeable outcome at the highest possible spread.
1. Reprice the line item. Tokens are COGS. Track cost per outcome, not cost per token, and expect the denominator to improve tenfold a year without anyone lifting a finger.
2. Demand layer attribution. Your AI bill should say which layer spent and which layer saved. A bill without a cache-hit line is a bill hiding your biggest discount.
3. Own one toll. Rent the power, rent the models if you like, but own your meter: the gateway, routing policy, and cache that govern your flow. In a deflationary input market, the meter is the asset.
The worked example
One AI app's P&L, before and after owning the meter
Theory earns its keep when it survives contact with an income statement, so here's the whole post compressed into one. Take an illustrative application at one million dollars of ARR, built the default way: every request goes to a frontier model, context ships untrimmed, nothing is cached, and the AI bill is one opaque line. Now give the same company the meter: a gateway that attributes spend, a router that sends mechanical steps to models a tenth the price, a cache that recycles the repeated 90 percent (the mechanics from our loop economics post, applied company-wide), and context discipline that stops paying to resend the same tokens. Same product. Same customers. Same revenue line.
Seventeen points of gross margin, no product change, no pricing change, no headcount change. That is what "own one toll" means in the currency a board understands. And note the second-order effect: the after-column company can now see, per layer, where every cent goes, which means next year's tenfold token deflation lands on its P&L automatically instead of vanishing into an opaque vendor line.
Today versus tomorrow
The smiling curve of inference, and how the smile moves
Economists who study manufacturing value chains draw the "smiling curve": value capture is high at the two ends of a chain, components on one side, brand and distribution on the other, and low in the middle where assembly happens [2]. The inference stack in 2026 wears exactly this smile: deep value at silicon, deep value at applications, a long thin valley across the middle. Overlay the forces from this post, deflation at the bottom, tolls in the middle, the convergence fight at the top, and you can draw where the smile moves.
The three migrations
Where the dollar rests tomorrow
Migration one: the bottom flattens. Silicon's margin is a supply story, and supply stories end. Fabs come online, custom accelerators fragment the buyer's market, and open engines erode the runtime moat one kernel at a time. Telecom ran this exact play: the companies that laid the fiber watched value flow over the top to the endpoints. The bottom of the stack doesn't go to zero; it goes to normal, and normal is not 78 points of gross margin.
Migration two: the middle tolls up. As spend explodes and per-token prices collapse, the questions that decide a company's AI economics all live in the sliver: which model gets this request (router), what does this run cost and who authorized it (gateway), what never needs computing again (cache), what does the model actually see (context). Every one of those decisions applies to 100 percent of flow, and the toll table above shows what history charges for that position. The monetization is already visible in embryo: routers pricing the spread between equivalent models, caches selling the absence of computation at margins software companies dream about, gateways charging like the interchange layer they are.
Migration three: the top deepens, and gets crowded. Distribution and proprietary context are the only assets in the stack that token deflation makes more valuable, because cheaper tokens mean everyone can afford intelligence, and the winner is whoever owns the place intelligence gets applied and the data it gets applied to. That's why the labs are climbing into this layer and why incumbents with existing customer relationships are suddenly dangerous again. Your context is the one layer nobody can rent, and after 2026, it's also the layer everybody is fighting for.
The twist ahead
When the buyer is an agent, attention stops working
Everything above assumes the oldest constant in commerce: a human at the top of the chain, choosing by habit, brand, and default. Now run the tape forward, because the fastest-growing buyer of inference is no longer a human. It's an agent, and agents buy differently. An agent has no home screen to be captured, no brand loyalty to be built, no Tuesday-morning habit. It evaluates every single call against three numbers: price, benchmark score, and latency. It will switch suppliers mid-task, forty times before lunch, without sentiment.
Think through what that does to the iron law. Attention-based power, the moat behind the App Store's 30 percent and the subscription app's 60 points, simply doesn't bind an agent. What binds an agent is whatever chooses on its behalf: the routing policy. In an agent-mediated economy, the router becomes the new distribution, the shelf every supplier must win placement on, per call, on merit. SEO becomes API documentation. Brand becomes benchmark score. The marketing budget becomes a price sheet. Suppliers of models and services will court routers the way consumer brands once courted supermarkets, and the toll table above gains its newest, strangest column. If you're building for that economy, the question isn't "how do I capture attention," it's "how do I win a policy," and those are entirely different games.
The global dollar
The tokenizer tax: the same intelligence costs more in some languages
One more secret hides inside the token itself, and it matters enormously for where the next billion dollars of AI spend comes from. Models don't price sentences; they price tokens, and tokenizers were overwhelmingly trained on English. The result, documented in tokenizer-fairness research, is that expressing the same meaning can require several times more tokens in other languages, with some scripts paying up to an order of magnitude more [4]. Many Indic languages sit toward the expensive end. Same question, same intelligence, same model: a materially higher bill, purely because of script.
For a CFO running multilingual products, the tokenizer tax is a real line item hiding inside the blended token price. For a builder, it's a market signal: the regions with the fastest-growing AI demand are often the ones paying the steepest per-thought prices, which makes every efficiency layer in the thin sliver, routing to token-efficient models, caching repeated context, trimming what's resent, worth more per dollar of flow there, not less. Affordable inference isn't one global price. It's an engineering outcome, achieved language by language, and the middle layers are where it gets achieved.
1. Pick the neutral ground. Build where both endpoints must cross and neither can own. Your moat is that you route to everyone, meter everyone, and favor no one; the fiercer the endpoint war, the more both sides need a Switzerland.
2. Price on share of savings. Charge a percentage of the bill you shrink, not a markup on the flow. It aligns you with the customer's CFO, and Jevons guarantees the base you're saving against keeps growing.
3. Publish the meter. Show every customer the receipt: per-layer attribution, cache-hit rates, routing decisions. Tolls run on trust, and in a market full of opaque AI bills, the company that shows its arithmetic wins the flow.
References
The economics this stands on
- Jevons, 1865. The Coal Question. The original efficiency paradox: making a resource cheaper to use multiplies its consumption. Substitute tokens for coal and read it again. full text
- Mudambi, 2008. Location, Control and Innovation in Knowledge-Intensive Industries (Journal of Economic Geography). The formalization of the smiling curve: why value concentrates at the ends of a chain and thins in the middle, until the middle finds a toll. doi:10.1093/jeg/lbn024
- Stanford HAI, 2025. The AI Index Report. The deflation data: the inference cost of a fixed capability tier fell roughly 280x between late 2022 and late 2024, the fastest input deflation on record. hai.stanford.edu/ai-index
- Petrov et al., 2023. Language Model Tokenizers Introduce Unfairness Between Languages (NeurIPS 23). The tokenizer tax quantified: the same content can cost several times more tokens across languages, up to an order of magnitude for some scripts. arXiv:2305.15425
Built on The 11 Layers of the AI Inference Stack. Related economics: The Cost of a Loop · Loop Engineering for CXOs. Written by Kashi and Rajan. Nothing here is investment advice; everything here is a map.
Test yourself
1. Where does most of an AI dollar come to rest today?
2. Why are the model labs racing into the application layer?
3. What does the cross-industry toll table imply for AI's middle layers?
FAQ
- How literal are the cents in the waterfall?
- Directionally solid, decimally illustrative. They're reconstructed from public reporting on application gross margins, lab inference margins, cloud economics, and chipmaker financials, then rounded to sum cleanly to 100. Your company's mix will differ. The claim we'd defend anywhere is the shape: three deep pockets, one astonishingly thin sliver, and the sliver touching everything.
- If everything in the middle is open source, how does anyone capture value there?
- The same way open source always monetizes flow: the software is free, the operated toll is not. Nobody pays for TCP/IP; everyone pays the payment network that runs on it. Expect the middle to monetize as managed meters, priced on flow governed or dollars saved, while the code itself stays open. Owning the standard and operating the toll are different businesses, and the second one is the good one.
- If the labs win the application layer, doesn't the middle die with the API?
- The opposite, structurally. A lab-owned router won't route to a competitor's cheaper model, which is precisely the router's value; a lab-owned gateway can't be trusted to meter its own bill. The middle layers are worth the most when they're neutral, the way card networks are worth the most because they clear everyone's transactions. Neutrality is a moat the incumbents at either end can't copy without ceasing to be themselves, and the more fiercely the endpoints fight, the more both need a Switzerland.
- Will strong applications just train their own models and cut out the labs?
- The strongest are already fine-tuning and some are training, which is the down-arrow in our convergence map. But full frontier training is a capital game with a tenfold-per-generation price tag, so for most apps the rational play is the router: rent whichever model is cheapest for each step, keep the spread. Which, note, routes even more strategic weight onto the neutral middle.
- What breaks this thesis?
- Two honest risks. If frontier capability keeps outrunning open weights indefinitely, the labs keep wholesale pricing power longer than we expect. And if one vertically integrated player wins distribution, models, and silicon simultaneously, the tolls get internalized instead of marketized, the way a company town replaces a market. We think both cut against the grain of every commoditizing input market in history, but this is a slope forecast, not a prophecy, and we've marked our curve accordingly.
- Doesn't the power and energy crunch change the map?
- It adds a twelfth participant rather than changing the law. Electricity is becoming the new silicon: a physical scarcity below Layer 11 that props up bottom-of-stack margins wherever grid capacity is rationed, and hands a structural edge to power-rich regions. But it follows the same arc every supply scarcity follows, capacity gets built, and it makes the efficiency layers more valuable in the meantime: a cache hit is a negawatt, and the cheapest data center is the request you never sent.
- Where does regulation move the dollar?
- Two predictable directions. Data residency and sovereign-AI rules pin flow to regional infrastructure, which redraws the cloud slice of the waterfall along borders. And any rule that mandates auditability, spend attribution, and human accountability for agent actions is, economically, a subsidy to the gateway layer: compliance turns the meter from a cost optimization into a legal requirement, which is historically how tolls become permanent.