Where the humans are
The LLM economy is settling into layers. At the top is question and answer: a person types, a model answers. Below that, agents do the actual work, calling tools, hitting APIs, talking to other agents. At the bottom, renderers turn all of it into something a human looks at.
Humans only touch the QA layer directly. Everything under it is bots talking to bots. That matters for advertising, because it tells you where an ad can honestly live. At the QA layer it looks like what ChatGPT already ships: a labelled sponsored unit under the answer. At the agent layer it looks different. An agent booking travel or picking a database is going to choose a vendor, and a vendor can pay to be one of the options, as long as it is disclosed and the organic options are still there untouched.
Both of those are real inventory. Today both belong to one seller.
ChatGPT started serving sponsored links under answers in February 2026: about $100M annualised inside two months, at roughly a $60 CPM.2 An LLM turn prices like paid search, because the person just told you what they want in their own words. Everyone else running a model gets none of that and pays for the GPUs anyway.
The YouTube argument
I keep coming back to YouTube. Nobody paid to learn coding there. The ads paid, and a generation of engineers came out of it. AdSense did the same thing for the written web in 2003:1 paste a snippet, get paid, no sales team. It did not make bloggers rich. It made publishing cheap enough that people kept doing it.
That is the trade I want for models. Not that providers get rich off ads. That ad revenue covers the compute, so the consumer side of a frontier model can be free, and the people who most need it, students, people in countries where $20 a month is real money, get the same model everyone else gets.
The provider side is the part that has to be motivated. If you run an open-weights model on your own hardware today there is no way for the value of a turn to reach you. Fix that, and serving open models stops being something you do at a loss until you get tired of it.
What it looks like, and how it gets paid
The unit is boring on purpose: the answer, a hard line, then a labelled block the model never saw. All the interesting parts are the things that cannot happen.
Kyoto ryokan, free cancellation
Traditional inns from $180 a night, cancel up to 24 hours before arrival.
See roomsUnderneath, one turn becomes one settled impression like this. The part I spent the most time on is the allocation: the exchange promises you a slice up front, and if you serve inside it you are paid. No more serving in good faith and finding out at settlement that the campaign ran out somewhere else.
Does it actually cover the compute?
I did not want to argue about this, so I worked it out from the hardware.
Cost. A 70B-class open model, fp8, batched, on an H100 at about $2 an hour, does roughly 3,000 output tokens a second. A turn of around 400 output tokens costs $0.000074 to decode; 1,200 input tokens add about a quarter of that in prefill. Call it $0.000093 a turn, nine cents per thousand.
Revenue. At a CPM of p, one ad every L
turns, and the node keeping 65 percent, you get 0.65p / 1000L
per turn. Set that equal to cost and the break-even CPM is
0.143 × L: $0.43 with an ad every three turns,
$1.42 every ten, $2.85 every twenty.
The market pays about $60. So this was never a question of whether ads can pay for inference. They pay for it a hundred times over. The question is whether the money can reach the person running the hardware without them having to build an ad business first.
The model author gets a cut too
Open-weights models are a public good with no revenue behind them. Every
node downstream captures the value. In this design the model's own manifest
at /.well-known/uap-model says what ads it allows and what
share it wants, and the exchange enforces it because the exchange is the
one holding the money. An author who wants none of it sets
permitted: false and no conformant exchange will clear an
auction on those weights. That is a veto and a royalty on traffic they
never see.
Plug in your own numbers:
The ad cannot touch the answer
An ad that pays per click, sitting next to text a model generates on the spot, is pressure to bend the text. Retail media already shows what happens: studies put 30 to 40 percent of sponsored-product sales down to purchases that would have happened anyway.4 Ipsos found about two thirds of US adults trust AI answers less when there are ads in them. A surface that bends its answers ends up at display CPMs.
So the protocol draws a line and makes it checkable, not just promised. The node hashes the answer and sends the digest to the exchange before the auction runs. At that moment no winner exists, so the answer cannot have depended on one. The composer that joins answer and ad is a plain function, so anyone can recompute the bytes:
from uap import verify_composition, verify_answer_commitment verify_composition(rendered, answer, decision) # (True, 'composition is exact')
The exchange also holds out a slice of requests: same path, no ad served, digest still reported. If the served answers drift from the held-out ones, something is bending them. No single receipt can show that; the holdout can.
Making it something providers can rely on
Motivation is not enough on its own. Providers will only plan around this if payment is predictable, so the parts I built first are the ones that make it stable.
| Built | What it guarantees |
|---|---|
| Signed allocations | Serve inside your slice and you are paid. Slices across all nodes cannot exceed the budget, so over-allocation is the exchange's loss. |
| Receipts recomputed, not trusted | Nonce, digest, price and pacing are all rederived at the exchange from what it issued. |
| Steward share enforced | The model author's cut cannot be reduced by the node or the exchange. |
| Buy side | Campaigns, creative review, forecasting, conversions, aggregate reporting. Agencies can spend without a private arrangement. |
| Integrity, provable | Answer digest committed before the auction; composition anyone can recompute; a holdout arm. |
Still open: the billing lifecycle, invoicing and disputes and make-goods. A production classifier; the one that ships fails closed on purpose, so it serves nothing until someone replaces it. And frequency capping across nodes, which I do not have a design for yet, because doing it network-wide needs an identifier the protocol forbids on purpose.
Nobody runs an exchange yet. Supply will not show up without demand and demand will not show up without supply, and a spec does not solve that. Somebody has to run all three roles in one narrow vertical until there is inventory worth bidding on. That is how every ad network started, including the closed one this is a response to. I would like that to be an open one.
If you want to attack something, attack the k-anonymity floor, whether the bundle fetch leaks anything about a node's audience, and whether the unattested tier has any honest demand. Those are the parts I am least sure about.
- Google, "Google Expands Advertising Monetization Program for Websites", 18 June 2003.
- Reported ChatGPT advertising revenue and CPM, February to April 2026. See Advertising on AI Platforms, 2026.
- Global digital ad spend 2026. See Digital Ad Spend Statistics 2026.
- Retail media sponsored-product cannibalisation estimates. See eMarketer, Amazon advertising FAQ.
- AdCP and IAB Tech Lab AAMP work above real-time bidding, at the campaign layer. UAP is the serve-time layer under both.