Blog · Book Review · Last reviewed July 2, 2026

The Age of Extraction and the Platform Tax

Tim Wu's The Age of Extraction names the economic habit that hides inside modern convenience. The platform begins as a route around friction, then becomes the route, then charges for passage through the world it helped reorganize.

The book belongs beside Wu's earlier work on information empires and attention capture, but its sharper target is the toll booth. Search, app stores, marketplaces, cloud services, social feeds, payment rails, model APIs, assistants, and default interfaces can all become places where a private intermediary takes money, data, attention, leverage, or institutional memory because everyone else has become dependent on the passage.

The Book

The Age of Extraction: How Tech Platforms Conquered the Economy and Threaten Our Future Prosperity was published by Knopf in November 2025. Penguin Random House lists the hardcover at 224 pages, and Amazon lists ISBN-10 0593321243 and ISBN-13 978-0593321249. Columbia Law's faculty books record places the book in law, economics, and technology, which is the right disciplinary neighborhood: it is not only a book about apps or advertising. It is a book about the institutional design of middlemen.

Wu's central move is to make extraction visible as a business relation. A platform can create value by helping people find one another, coordinate, transact, publish, navigate, or compute. The trouble begins when the same platform becomes unavoidable and then uses that position to take a larger share of the exchange than competition, consent, or public usefulness can justify.

That argument extends the arc of The Master Switch and The Attention Merchants. First, communications systems consolidate. Then attention becomes an industrial object. Then the consolidated attention and infrastructure layer becomes a rent point. The platform does not merely persuade people to look. It stands between other people and the things they need to do.

The book is short, pointed, and intentionally public-facing. It is less a complete theory of capitalism than a warning about a particular institutional pattern: the useful intermediary hardens into a private checkpoint, and the checkpoint normalizes the toll.

What Extraction Means

Extraction is not just profit. A firm can earn profit by building a better product, reducing cost, expanding supply, or solving a coordination problem. Wu's target is narrower and more corrosive: the power to take from a transaction because one controls the route through which the transaction must pass.

That route can be economic, informational, social, or cognitive. A marketplace can extract from sellers through fees, ranking dependence, logistics terms, advertising pressure, and the threat of invisibility. A social platform can extract from users through attention, behavioral data, identity infrastructure, and social dependency. A mobile operating system can extract from developers through app-store rules, payment requirements, search defaults, and review gates. A cloud provider can extract from institutions by making compute, storage, identity, logs, and deployment habits difficult to move.

The useful concept here is the platform tax. It is the bundle of fees, defaults, ranking rules, data access limits, self-preferencing, switching costs, API restrictions, advertising requirements, and opaque enforcement that attaches to dependency. Some of it appears as money. Some of it appears as data. Some appears as lost bargaining power. Some appears as time spent complying with an interface that cannot be negotiated with.

This is why the word "platform" matters. A platform is not merely a company with users. It is a company that other people build on, sell through, communicate inside, learn from, or rely on as infrastructure. The deeper the dependency, the easier it is to describe the tax as normal operating cost. What looked like convenience from the front becomes governance from the back.

The AI-Era Platform

The Age of Extraction matters in 2026 because generative AI and predictive social data can make extraction feel like assistance. The toll booth no longer has to look like a checkout page or an advertising auction. It can look like a helpful answer, a default assistant, an automated workbench, a model API, a retrieval layer, or a routing decision made before the user sees the alternatives.

AI systems add three new extraction surfaces. The first is compute and model access. If a small number of firms control the practical route to frontier models, chips, cloud infrastructure, safety tooling, evaluation services, and distribution, they can charge not only for software but for participation in the new work environment. The second is data and feedback capture. Prompts, corrections, clicks, files, workflows, and institutional edge cases become training or product signals. The third is interface control. The assistant can decide which sources appear, which vendors are called, which tools are offered, which answers feel complete, and which routes are never shown.

This is platform power translated into cognition. An AI search system can become the front door to public knowledge. A coding assistant can become the front door to software work. A procurement assistant can become the front door to vendors. A workplace assistant can become the front door to memory, policy, and administrative action. Each of these products may be genuinely useful. Usefulness is exactly how dependency forms.

The competition-authority joint statement on generative AI is useful background because it names concentrated control over inputs such as chips, compute, data, and expertise as a competitive concern. The statement's principles of fair dealing, interoperability, and choice map cleanly onto Wu's warning. Extraction thrives when fair dealing is weak, interoperability is blocked, and user choice is mostly decorative.

The sharper AI-era question is not whether a system is clever. It is whether the system is becoming the required route. When a model, assistant, app store, cloud account, or search default becomes the route, its governance is no longer a private product-detail issue. It is a question about who can act, publish, discover, buy, sell, remember, and be found.

Governance and Remedies

The practical contribution of Wu's frame is that it turns platform power into an audit object. Institutions should keep a platform-tax register: every material fee, default, ranking rule, payment requirement, data access restriction, export limit, self-preferencing risk, API dependency, content-training claim, appeal process, and switching barrier that attaches to a platform relationship.

That register should be paired with a switch register. For each critical platform, the institution should know what it would take to leave, interoperate, mirror, export, federate, or run a fallback route. The answer will often be uncomfortable. That is the point. Dependency is easiest to measure before crisis, policy change, price increase, account suspension, or model deprecation.

For AI systems, a serious register should include model provider, cloud provider, retrieval indexes, data retention settings, training-use terms, logging rules, prompt and file exportability, evaluation rights, safety-review routes, human appeal, incident records, rate limits, identity binding, and downstream tools the assistant can operate. If a public institution relies on an AI intermediary, the public should not have to guess how records, answers, decisions, or referrals are being routed.

Antimonopoly policy is part of the answer, but not the whole answer. Competition cases can challenge exclusion, tying, self-preferencing, acquisitions, and monopoly maintenance. Procurement rules can preserve portability. Public-interest duties can require notice, appeal, logs, and nondiscriminatory access. Technical standards can make exit and interoperability real. Labor and creator rules can limit the conversion of human work into training fuel without consent, compensation, or refusal rights.

Wu's warning is strongest when read as institutional hygiene. Do not let a convenience become an unexamined dependency. Do not treat a smooth interface as evidence of fair dealing. Do not confuse a platform's growth with the public's prosperity. Build exit before exit is needed.

Limits

The word "extraction" can do too much work if it is not disciplined. Not every commission is abusive. Not every integration is a trap. Not every default is illegitimate. Platforms can reduce fraud, improve discovery, lower transaction costs, standardize payments, protect security, and make small actors visible. A useful critique has to distinguish value creation from toll collection.

The evidence layer therefore matters. A platform-tax claim should identify the fee, rule, dependency, affected group, counterfactual route, switching cost, and competitive harm. Without that discipline, "extraction" becomes a mood instead of an analysis.

The book also leaves room for companion work on labor, the environment, the global political economy of chips and data centers, and the public-sector procurement problem. The extraction of money, data, and attention is not the only extraction attached to platform capitalism. Energy, minerals, warehouse labor, content moderation, training-data labor, and public subsidy all belong in the larger account.

Those limits do not weaken the book's core warning. They make the warning more useful. The platform tax is one layer of a larger system, and it is a layer institutions can start documenting now.

What This Changes

For this archive, The Age of Extraction supplies a vocabulary for the AI middleman. It connects monopoly, attention, data capture, cloud dependency, answer engines, app stores, assistants, and creator economics without pretending they are the same industry.

The shared pattern is route control. A platform captures a path between people and their work, audience, knowledge, customers, tools, or memory. Then it learns from the traffic. Then it optimizes the route for its own durable advantage. Then everyone downstream starts adapting to the route as if it were nature.

The AI version will be subtler because it will often present itself as personalization, summarization, safety, productivity, or care. The assistant that saves time can also become the agent that decides what is worth seeing. The search answer that reduces friction can also erase the source. The model API that lets a startup build can also make the startup a tenant. The cloud convenience that helps an agency move quickly can also make public administration dependent on private terms.

Wu's book does not ask readers to reject every platform. It asks them to stop being naive about the moment when helpful infrastructure becomes a private gate. That is the right question for AI governance: where is the gate, who controls it, what does passage cost, what records are kept, what alternatives exist, and who can contest the toll?

Source Discipline

This review uses Penguin Random House, Amazon, and Columbia Law for bibliographic metadata and publisher framing; Brookings for public discussion context around the book and platform concentration; the joint competition-authority statement for AI input-concentration and interoperability concerns; the FTC AI partnerships report for recent regulatory attention to AI investments and partnerships; and the DOJ Google case page for the broader antitrust backdrop around search and platform power.

Claims about the book's argument are interpretive, anchored in publisher and catalog descriptions plus Wu's established public work. Claims about law and competition policy are jurisdiction-specific and should not be treated as legal advice. The review uses "platform tax" as an analytic label for observable fees, defaults, data terms, ranking rules, and switching costs, not as a statutory term.

Sources

Book links are paid affiliate links. As an Amazon Associate I earn from qualifying purchases.

Buy on Amazon Browse Books

Return to Blog · Return to Books