Aggregation Theory is a framework, advanced by Ben Thompson of Stratechery in July 2015, describing how the Internet shifts economic power away from incumbents that control the distribution of scarce supply and toward "aggregators" that control the demand-side relationship with users. Thompson refined and formalized the framework in Defining Aggregators (September 2017). It has since been applied as an analytical lens to assess whether individual AI companies are aggregators, integrators, or a category Thompson has not formally named.
Core claim
In the pre-Internet era, value chains had three components — suppliers, distributors, and consumers — and outsize profits accrued to whoever integrated backward into supply by controlling distribution. Thompson's examples include newspapers, networks, taxi companies, hoteliers, and book publishers.
Thompson argues the Internet broke this dynamic in two steps:
- Distribution of digital goods became free, neutralizing the incumbent's primary moat.
- Transaction costs collapsed to zero, making it viable for a new kind of entity, the aggregator, to integrate forward into the consumer relationship at scale.
On this account, value shifts away from incumbents (who control distribution of scarce supply) and toward aggregators (who control demand for abundant supply): suppliers commoditize and aggregators win.
The 3-level classification
In Defining Aggregators, Thompson defines an aggregator by three characteristics: (1) a direct relationship with users; (2) zero marginal cost to serve a user; and (3) a demand-driven multi-sided network with decreasing customer-acquisition costs.
Aggregators are classified into levels by their relationship to suppliers:
- Level 1 — Supply Acquisition. The aggregator pays for or acquires supply (Netflix). This level is slower-growing and more vulnerable to deeper-pocketed competitors.
- Level 2 — Supply Transaction Costs. The aggregator does not own supply but bears onboarding and transaction friction (Uber, Airbnb in some jurisdictions).
- Level 3 — Zero Supply Costs. Suppliers come for free, often actively competing for placement (Google, Facebook).
Super-Aggregators (Facebook, Google) operate three-sided markets — users, suppliers, and advertisers — with zero marginal cost on all three sides.
Application to AI companies
Aggregation Theory has been used to assess whether a given AI company holds a winning strategic position:
- OpenAI is characterized as an accidental Level 3 or Super-Aggregator candidate via ChatGPT, though Thompson argues it has moved too slowly to monetize through advertising and risks losing scale to Google and Microsoft (per AI Promise and Chip Precariousness). Benedict Evans disputes the aggregator reading outright in How will OpenAI compete?: ChatGPT's 800–900 million users are weekly actives with shallow engagement rather than an aggregator's daily relationship, memory and similar features are "stickiness, not a network effect," and running infrastructure confers no leverage over products further up the stack, since users of an application do not know which model it calls. Evans also rejects the claim that owning agent and commerce protocols would supply the missing network effect, on the grounds of the widget fallacy and low developer switching costs.
- Google and Meta are already Super-Aggregators leveraging existing user relationships into AI distribution; FANG Playbook argues they will subsume rather than be disrupted by AI.
- Anthropic is described as a capability-bet integrator rather than an aggregator, relying on enterprise-API revenue rather than a consumer-relationship moat (per Anthropic and Alignment and Mythos, Muse, and the Opportunity Cost of Compute).
- DeepSeek, Llama, and gpt-oss are open-weight models that commoditize the supply side, putting pricing pressure on closed-model APIs and accelerating the shift of value toward whoever owns the consumer relationship.
- Cursor/Anysphere, Perplexity, and Replit are Level 1-or-2 aggregator candidates in vertical AI applications, dependent on whether they can keep their direct user relationship as base-model providers commoditize.
The framework also informs the AI-and-antitrust debate (AI Antitrust): on Thompson's account, aggregators win not by abusive practices but by offering a better user experience, which complicates consumer-welfare antitrust analysis.
Debates and tensions
- Thompson's view that aggregators are inevitable — that the only choice is which one wins, not whether to allow them — is in tension with the FTC under Lina Khan and the Khan-school antitrust literature, which argues that structural-separation remedies could prevent aggregator consolidation.
- The framework does not cleanly handle vertically integrated AI labs that both train models and operate consumer products, which are neither a pure aggregator nor a pure incumbent. OpenAI, Anthropic, and Google increasingly resemble a new category Thompson has not yet formally named.
- Open-weight models partially break the supply-commoditization step: when the supply (model weights) is a free good available to anyone, the aggregator's leverage weakens. Thompson discusses this in AI Promise and Chip Precariousness §"Goldilocks".
Relationships
- depends-on: Ben Thompson — originator
- supports: Aggregation Theory — Ben Thompson (Stratechery, 2015) — canonical 2015 essay
- supports: Defining Aggregators — Ben Thompson (Stratechery, 2017) — 3-level classification
- supports: The FANG Playbook — Ben Thompson (Stratechery, 2016) — applied to Big Tech
- related: AI Bubble Debate — aggregation dynamics inform whether current AI capex is sustainable
- related: Open-Source AI / Open-Weight Models — open weights complicate the supply-commoditization step
- related: AI Antitrust — Aggregation Theory provides the conceptual frame for AI-era antitrust analysis
- contradicts: Christensen-style low-end disruption narrative — see What Clayton Christensen Got Wrong — Ben Thompson (Stratechery, 2013)