---
title: "When Models Become Commodities, Where Does the Money Flow? — Moats in the AI Value Chain"
url: https://blog.tyrano.dev/en/when-models-become-commodities-where-does-the-money-flow-moats-in-the-ai-value-chain
lang: en
site: Tech AI News - 테카이
category: "AI Industry Analysis"
tags: ["AI","artificial intelligence","AI value chain","moat","commoditization","token economy","LLM","business model"]
published: 2026-10-26T07:11:24.275Z
updated: 2026-10-26T12:29:24.463Z
sources:
  - https://www.normaltech.ai/p/up-the-stack-how-ais-escape-from
  - https://a16z.com/2023/01/19/who-owns-the-generative-ai-platform/
  - https://clariceqiu.substack.com/p/the-year-the-commodity-layer-ate
  - https://advisoranalyst.com/2026/09/22/the-intelligence-layer-mckinsey-maps-14-technologies-reshaping-the-economy-in-2026.html/
  - https://new.philippdubach.com/posts/is-ai-really-eating-the-world-1/2/index.md
---

# When Models Become Commodities, Where Does the Money Flow? — Moats in the AI Value Chain

> If two companies use the same model but one makes money and the other doesn't, the difference isn't the model — it's where each one sits, above or below it, in the value chain.

---

## Background — why this topic, why now

AI model prices keep falling. According to an analysis tracking API pricing ([Philipp Dubach, May 2026](https://new.philippdubach.com/posts/is-ai-really-eating-the-world-1/2/index.md)), per-token prices for comparable-performance models have dropped by as much as 97% since launch. Taken on its own, this trend points to the natural conclusion that "models are becoming a commodity" — a good whose quality differences vanish, leaving price as the only axis of competition.

The problem comes next. Just because models become commoditized doesn't mean the applications built on top of them automatically make money. This piece splits the AI value chain into model, application, and operational-residue segments (the data and workflow dependencies that accumulate per customer) and looks at where real defensibility — a "moat," a structural advantage competitors can't easily cross — actually forms.

## Key data & current state

- **API price decline:** Up to 97% for comparable-performance models (Philipp Dubach, compiled May 2026)
- **Anthropic annual recurring revenue (ARR):** Roughly $9 billion at the end of 2025 → more than $44 billion by mid-2026, with inference gross margin rising from about 38% to over 70% (SemiAnalysis estimate, [as cited by Clarice Qiu](https://clariceqiu.substack.com/p/the-year-the-commodity-layer-ate); not audited financial statements)
- **Companies' perceived AI payoff:** 89% of organizations use AI regularly in at least one function, but only 37% report a positive perceived impact on revenue (McKinsey 2026 survey, [as cited by AdvisorAnalyst](https://advisoranalyst.com/2026/09/22/the-intelligence-layer-mckinsey-maps-14-technologies-reshaping-the-economy-in-2026.html/))

> Reading: Token prices are falling while model companies' revenue is surging, and even though companies use AI widely, only about a third feel a revenue impact. That's a sign that "value creation" and "value capture" have come apart. The total value AI generates keeps growing, but how much of it each layer gets to keep varies widely.

## In-depth analysis

### The model layer — commoditization hits "easy work" first

The commoditization argument assumes performance convergence and low switching costs. That premise holds for general-purpose tasks. But [Clarice Qiu's rebuttal](https://clariceqiu.substack.com/p/the-year-the-commodity-layer-ate) points out that this view only looks at the "cost of average intelligence." Easy tasks commoditize quickly, but enterprise willingness to pay concentrates on harder tasks. For work like coding agents that replace human labor, the price anchor isn't the model's price — it's the labor cost being displaced. Analysis: the model layer doesn't seem to be flattening into "all models are the same," but rather into "easy tasks converge, hard tasks still differ."

### The application layer — thin wrappers get squeezed out first

Services that merely wrap a model API — "thin wrappers" — absorb every price and performance shift at the model layer directly. As inference costs fall, apps that had differentiated themselves with outcome-based pricing shifting back toward flat-rate or seat-based pricing is also read as a result of this pressure. An a16z analysis of the generative AI market from early 2023 ([a16z, "Who Owns the Generative AI Platform?"](https://a16z.com/2023/01/19/who-owns-the-generative-ai-platform/)) concluded at the time that there was "no discernible systematic moat." Nearly three years later, that picture hasn't changed much overall — except that the gap has started to widen depending on "what kind" of application it is.

### Operational residue — neither model nor app, but what accumulates

An analysis from [normaltech.ai](https://www.normaltech.ai/p/up-the-stack-how-ais-escape-from) argues that the moat comes less from the layer itself than from what "accumulates" on top of it. Examples include memory, documents, and workflows that build up per customer (an embedding moat), multi-year contracts and bundling (a commercial moat), and the effect of users growing accustomed to how a tool works (a behavioral moat). Because the cost of migrating all that accumulated data and familiar workflow is far higher than the cost of switching models, these moats don't disappear easily even when the underlying model changes. Analysis: the flatter the model layer gets, the sharper, paradoxically, the advantage of whoever has accumulated the most operational residue may become.

## Implications for Korea

Most companies and institutions in Korea sit at the application or internal-tooling layer, not the model or infrastructure layer. Applying this analysis, what matters more than "which model you use" becomes "how much of the data and business processes around the model you keep locked into your own environment." That means that while keeping the option to switch model vendors open, designing AI workflows to accumulate assets that can't be easily moved — internal documents, work history, approval flows — should get more weight than model selection itself. The same logic applies to fields like content and DAM (digital asset management), where AI gets layered onto an enterprise's content pipeline. Analysis: the defensibility of such tools is more likely to come from how deeply they're woven into existing content assets, metadata, and approval processes than from generation quality alone.

## Outlook & variables to watch

- **Upside variable:** If inference costs keep falling, the range of tasks AI can be applied to — including ones previously ruled out on cost — could widen, growing the overall application-layer market.
- **Risk variable:** As operational-residue-based moats strengthen, there's a risk that choice narrows toward a handful of platforms that accumulated data first. normaltech.ai suggests that interoperability and switching-cost standards are needed before this kind of lock-in hardens.
- **Checkpoint:** Distinguishing whether a given service's revenue growth comes from "high-volume processing of easy tasks" or from "the substitution value of hard tasks" is useful for reading how value capture is actually structured.

## Conclusion

- Price competition at the model layer is real, but it's most pronounced in "easy tasks," while a gap still remains in "hard tasks."
- The application layer as a whole isn't automatically advantaged — services that merely wrap a model thinly are actually the first to feel the squeeze.
- Real defensibility tends to show up more clearly in the data and workflow dependencies that accumulate per customer on top of a layer than in the layer itself.
