The seat that actually wins in the AI era: the token path and the paradox of value capture

The seat that captures value in the AI era is neither the model nor the app but a defensible token path. In 2026 a16z proposed "being in the token path" as the new rule for picking winners, and the same year Benedict Evans warned that even there the productivity gain gets competed away. ASAP synthesizes the two insights into what actually wins in the AI era.

The question moved from "what do you own" to "where do you sit"

David George of a16z says the rule for picking winners changed in 2026. The market grows larger, but predicting who captures that value has gotten harder. Instead of oscillating between "model or app," the number one screen now is whether a company sits in the token path.

The token path is the channel where LLM inference, from token input to token output, actually happens. Sitting in that channel brings two things. First, it captures value automatically from the flow as AI usage explodes. Second, it is hard to remove because it holds the workflow.

Why this frame matters

Read it as a signal that the center of gravity in investment judgment has shifted from owning assets to occupying a position. What you built matters less than where that output sits in the flow of usage. Whether you own a model, or whether the app looks impressive, becomes a secondary question; the primary question is whether you sit in the channel that usage passes through when it explodes. Position itself becomes the basis for charging.

Why position alone is not enough

Benedict Evans warns that value can be competed away even when you sit in the token path. If a discounted-cash-flow analysis drops from a week to 10 seconds, you do 50 times more of them, but you cannot charge more. A feature anyone can build becomes a competitive necessity, the productivity gain is offset, and the surplus flows to customers rather than to vendor margins.

This is where the two insights complete each other. If a16z asks "where to sit," Evans asks "whether that seat is distinguishable from anyone else's." Sitting in the flow is worthless if the flow is substitutable, because pricing power converges to zero. The answer narrows to a defensible token path. Merely sitting in the flow is not enough; you must hold the seat with proprietary data, a hard-to-break workflow, and accumulating context. Only then can you avoid being copied when model companies leg up into applications.

A local lens: what you should build

What you need is not a new foundation model but an agent that holds a proprietary workflow. Training a model is a capital-intensive game for a few labs, and the token path does not mean owning a model. You use models via API and layer your own context, memory, tools, and workflow on top to hold a flow that is hard to copy. Cursor, through which every line of a developer's generated code passes, is one example of holding that seat.

Applied to the Korean market, the implication is clear. Because foundation-model competition is, by sheer capital scale, a game for a few labs, the battleground for most local teams is a defensible token path that digs deep into a specific workflow. Local-specific context such as regulation, language, and industry-specific documents can itself become a hard-to-copy moat. This is an interpretation drawn from the frame, however, and which flows actually hold defensive power is something each team must verify.

Limits and counterarguments

The frame is powerful but not a cure-all. Judging something "defensible" tends to be retrospective, and there is no guarantee that today's proprietary workflow will not degrade into tomorrow's competitive necessity. Evans's warning applies just as much to a seat you believed was defensible. A case like Cursor is only an example of occupying a position, not proof that the seat is permanent. It is safer to treat the frame as a questionnaire rather than a conclusion, using it to repeatedly ask "am I in the flow, and is that seat truly hard to copy?"

Source: a16z, "The New Rule for Picking AI Winners" (2026-05-29); a16z, "The Economics of AI Usage and What's Next For SaaS" (Benedict Evans, 2026-06-08).