If you build AI apps, invest in AI distribution, or keep asking “where exactly is my moat” — this is the new battlefield of AI-era distribution: it no longer fights for your attention; it fights for the agent’s power to choose.

The most profitable playbook of the past twenty years is Aggregation Theory: whoever directly owns the user can commoditize the suppliers, force them to undercut each other, and skim the profit. Google turned websites into interchangeable blue links; Amazon turned merchants into an interchangeable shelf.

Put “the model” in the supplier column and the entire AI downstream welds onto this playbook. AI makes it fiercer — intelligence is deflating (capability matched by open source month by month, prices collapsing), so the model layer is becoming exactly the kind of supply an aggregator dreams of: powerful, cheap, interchangeable, and getting cheaper. Cursor can squeeze its model suppliers not because Cursor is strong, but because the model category is being commoditized while Cursor holds the developers.

But here’s the turn most people miss —

The Object Changed: From “Attention” to “Intent Routing”

Classic aggregators aggregate human attention: Google aggregates search intent, Meta aggregates eyeballs, then sells the attention to advertisers. A trillion-dollar model built on “people will look, people will click.”

Agents pull that premise out. When an AI agent searches, compares, and orders on your behalf, attention as a sellable scarcity starts to dissolve (brands don’t die, but they shift from “persuading people” to “shaping the agent’s ranking and trust signals”). The object of aggregation moves from human attention to the routing of intent: a request comes in, and who decides which model, which tool, which supplier it goes to.

That is the real aggregation point of the AI era. Its moat isn’t brand and habit — it’s workflow embedding + accumulated context. The more an agent knows your context, the less anyone can replace it to route your intent.

Three Aggregation Points, Devouring Top-Down

Intent crosses three layers from origin to execution, each with a contender for aggregator:

• OS / hardware entry: closest to raw intent, able to set itself as the default agent. The highest point — and what model companies fear most: if a phone maker privatizes the system-level AI entry, even a dominant ChatGPT gets demoted to supplier.
• Super-app / general assistant: how model companies attack the aggregation point themselves, betting they can weld “ask AI” to their brand before the OS reacts.
• Vertical agents: don’t fight for general intent, just seize the routing power of one high-value scenario. Narrow, but deep.

The three relate by top-down devouring: an OS can demote a super-app to an icon, a super-app can absorb a vertical agent into a plugin. But with a boundary — this devouring is strongest in consumer/light-task scenarios; in enterprise, high-compliance, deep-process scenarios, the vertical layer holds permissions, audit trails, and procurement relationships, and may instead lock up the entry from below.

Three closing lines. One: to value an “AI app,” first locate which of the three layers it sits in, and whether the layer above has the motive and ability to swallow it. Two: most overvalued is the app with many users but no intent routing (attention aggregation depreciates in the agent era); most undervalued is the vertical agent that’s narrow but owns intent routing + accumulated context. Three: watch one gauge — the standardization/portability of intent routing; once it’s punched through by tech or regulation, the entire downstream aggregation rent collapses together.

Downstream is the good end, but downstream has its own upstream — the closer to intent, the more you win; and the layer closest to intent is one the model companies may not get to keep.

Your product or position — does it hold the intent entry, or just a pile of users? Comments open.

Sources (framework argument, June 2026): Aggregation Theory (Ben Thompson’s classic framework); model-layer commoditization mechanics in Chapters 1 & 6 (intelligence deflation, open-source catch-up; data from Stanford AI Index, SemiAnalysis).

— From Chapter 8 of a book in progress, working title The Deflation Sandwich

#AI #AIIndustry #DeflationSandwich