Fragmentation

Chapter 1: The Platform Bargain

In 2004, a 19-year-old computer science student at Harvard named Mark Zuckerberg launched a website called "TheFacebook" from his dorm room. It was one of thousands of internet projects launched that year, most of them forgotten. There was nothing obviously revolutionary about it. The site did not invent social networking—Friendster, Orkut, and MySpace already existed. It did not pioneer the concept of sharing personal information online. It did not solve a technical problem that was otherwise unsolvable. What it did was implement a specific economic model with ruthless clarity, and that model would reshape how billions of people experienced information, connection, and knowledge for the next two decades.

The model was simple: offer a free service that connected people to their friends and the information they cared about. Extract detailed data about users' behavior, preferences, and social graphs. Use that data to sell advertising with unprecedented precision. Grow the user base constantly, because growth made the advertising more valuable. Never charge users directly, because charging would limit growth. Never build interoperability with competitors, because walled gardens were more profitable.

This was not a novel insight. Advertising-supported media had existed for centuries. What was novel was the scale and scope at which this model could operate, enabled by the internet's technical architecture and the declining cost of computing. A newspaper or television station had to serve a geographic market and could extract only limited information about its audience. Facebook could serve the entire world and extract information about what every single user read, clicked, liked, searched for, and lingered over—second by second, across billions of people, indefinitely.

By 2010, less than six years after launch, Facebook had half a billion users. By 2020, it had nearly three billion. Google, which had implemented a similar model for search—free service, behavioral data extraction, targeted advertising—had captured search. YouTube captured video. Instagram captured photo sharing. Twitter captured short-form public conversation. Amazon captured e-commerce and cloud computing. These platforms did not emerge from some central plan. They were not designed by a consortium of internet architects with a singular vision. Instead, they resulted from a straightforward economic logic: if you could give away a service for free, extract data about millions of people, and monetize that data by selling access to advertisers, you could build a business of almost unlimited scale.

The model worked. It worked so well that it came to define what "the internet" meant to most people. When a teenager in Indonesia, a farmer in Brazil, or a student in Nigeria thought of "the internet," they often meant Facebook, WhatsApp, or YouTube—the platforms that were free to use, where their friends were, where the content was. The underlying infrastructure—TCP/IP, DNS, HTTP—was invisible to them. What was visible was the platform. And the platform had a deal to offer: participate in our network, give us your data, accept our advertisements, and in return, you get to connect with everyone you know and discover content algorithmically tailored to your interests.

This bargain felt, for a long time, like an exceptional offer.

To understand why requires stepping back to the 1990s and early 2000s, before the platform model had fully consolidated. The early internet was fragmented. If you wanted to find a website, you used different search engines—AltaVista, Lycos, Yahoo. If you wanted to send email, you chose an ISP that came with email service, and your address was locked to that ISP. If you wanted to chat, you used AOL Instant Messenger or ICQ, but people on AOL couldn't talk to people on ICQ. If you wanted to share files, you used your own server or relied on someone else's limited infrastructure. The early web was a patchwork of incompatible systems and limited reach.

More fundamentally, the early internet had a discovery problem. If you built a website, almost no one would find it. There were millions of websites, and no systematic way to navigate them. Search existed, but search required knowing what you were looking for. If you wanted to discover something new—a video you'd enjoy, an article relevant to your interests, a person you might want to follow—there was no infrastructure for that. Every media company faced the same challenge: How do you help people find what they want when there are billions of possible items to choose from?

The answer was algorithmic curation. Google's PageRank algorithm ranked search results not by keyword matching but by the structure of the web itself—by counting links, determining authority, and inferring relevance from how the web had organically connected itself. It transformed search from a keyword-matching tool into a relevance engine. Facebook's News Feed, launched in 2006, applied the same logic to social media: instead of showing you posts chronologically from everyone you followed, it showed you the posts most likely to engage you, based on your past behavior and the behavior of people like you. YouTube's recommendation algorithm did the same for video.

These algorithms required data to work. The more data they had—about what you clicked, how long you stayed, what you liked, what made you scroll—the better they could predict what would keep you engaged. The more engaged you stayed, the more opportunities there were to show you advertisements. The cycle was self-reinforcing: better data led to better recommendations, which led to more engagement, which led to more data.

For users, this created something new: a genuinely useful service that required no payment. Google solved the problem of finding information on the internet. Facebook solved the problem of maintaining connection with people you knew and discovering content from strangers. YouTube solved the problem of finding video. These were not trivial problems. Before these platforms existed, these tasks required significant effort and knowledge. After they existed, they were free, frictionless, and universal.

But the bargain extended beyond mere utility. The platform model also created what economists call "network effects"—the phenomenon where a service becomes more valuable as more people use it. A social network with a million users is useful. With a billion users, it becomes essential, because that's where almost everyone you know is. A search engine with more pages indexed is better. The dominant search engine with the most pages indexed, the most queries, and the most data about what people find relevant is vastly better. An e-commerce platform with one million products is convenient. One with hundreds of millions of products, processed through algorithmic recommendation and expedited shipping, is transformative.

These network effects created a winner-take-most dynamic in each category. Once Facebook became the dominant social network, there was little reason to split your time between multiple social networks—your friends were on Facebook, and fragmenting your attention reduced the value of the service. Once Google became the dominant search engine, competitors offering marginally better results were irrelevant, because search results were good enough and the network effect of scale made them better. The platforms competed brutally in the early stages, but once one achieved dominance, competition largely ceased.

From a user's perspective, this consolidation was not obviously bad. It meant one platform where your friends gathered. One search engine. One place to watch video. It was actually more convenient than the fragmented web it replaced. It was more useful because the scale enabled better algorithmic recommendations. It was more connected because network effects meant almost everyone was on the same platform. The bargain seemed good: give us your data, accept advertisements, and in return, you get access to billions of people and content tailored to your interests.

The platforms themselves recognized the strength of this position. By the early 2010s, the largest platforms had begun to move beyond their original categories. Facebook launched its own messaging service and bought Instagram and WhatsApp, integrating them into its ecosystem. Google bought YouTube and Android, making its services unavoidable on smartphones. Amazon moved from e-commerce into cloud computing, entertainment, grocery delivery, and health care. These weren't just expansions—they were reinforcements of the core model. More services meant more data. More data meant better algorithms. Better algorithms meant more engagement. More engagement meant more advertising revenue.

There was something almost elegant about the model from a business logic perspective. You didn't need to charge users because users were not the customer—they were the product. Or more precisely, they were the asset. The real customers were advertisers, who wanted access to billions of people at scale, segmented into micro-targeted audiences based on their detailed behavioral profiles. The platforms were the gatekeepers between users and advertisers, and that gatekeeper position was extraordinarily profitable.

By 2020, five companies—Google, Facebook (Meta), Amazon, Apple, and Microsoft—controlled vast portions of digital life. Google controlled search, email, cloud storage, and video. Facebook controlled social networking, messaging, photo sharing, and increasingly, commerce. Amazon controlled e-commerce, cloud computing, and digital streaming. Apple controlled smartphones and increasingly, digital wallets and services. Microsoft controlled productivity software and business cloud infrastructure. These were not separate silos—they overlapped, reinforced each other, and together they had integrated themselves into nearly every consequential digital transaction.

And yet, this concentration did produce something genuine that previous media eras had not: a period of remarkable connection and democratization of voice.

A teenager with a smartphone in rural India could reach a global audience on YouTube or TikTok. A journalist in an authoritarian country could publish to millions on Twitter without needing institutional support. A person could build a business through e-commerce on Amazon without building their own logistics infrastructure. A student could access world-class educational content on YouTube for free. A researcher in a small university could publish findings on Google Scholar and be discovered by anyone in the world. For a specific window—roughly 2005 to 2020—the platform model created what the media scholar Clay Shirky called "the era of cheap and universal participation."

The platforms were profitable, consolidated, and essentially unaccountable to anyone but their shareholders. But they were also, for a moment, genuinely connecting and democratizing in their effects.

This was the Platform Bargain: trade your behavioral data and attention for access to a universally connected network, algorithmic discovery, and services of remarkable usefulness, all at zero monetary cost. The trade-off was not symmetric—the platforms benefited more than users—but it was not abusive in its basic logic. The platforms were extracting value from user data and attention, yes, but they were also providing value back, in the form of services that demonstrably improved people's lives. A farmer in Kenya could access weather information and market prices through Google Search. A small business could reach customers through Facebook. A researcher could discover papers through Google Scholar. A person isolated during a pandemic could maintain connection through video calls on platforms like WhatsApp and Zoom.

The political economy of this arrangement was contested from the beginning. Regulators worried about monopolistic practices. Privacy advocates worried about data extraction. Researchers documented filter bubbles and algorithmic amplification of extreme content. Labor advocates noted that platform work was precarious and exploitative. These criticisms were not wrong. But they were occurring against the backdrop of a system that was, for much of its duration, genuinely useful and genuinely connecting in ways that previous media technologies had not been.

The stability of the Platform Bargain depended on several conditions holding true. First, it required that the primary value of the platforms came from their size, ubiquity, and algorithmic curation of user-generated or publicly available content. If the real value started coming from something else—from proprietary data, from domain-specific expertise, from integration with specialized workflows—the logic of the bargain would change. Second, it required that the cost of providing these services remained low relative to advertising revenue. If the cost of computing, data storage, or moderation rose significantly, the economics would shift. Third, it required that competition remained structured around who could build the largest network and extract the most behavioral data. If competition shifted to something else—to who could build the most sophisticated proprietary systems, or who could offer the most specialized integration—the entire logic would invert.

Beginning around 2022, all three of these conditions started to break down simultaneously.

The immediate catalyst was artificial intelligence. Not the narrow AI of recommendation algorithms, which platforms had been using for a decade, but large language models and diffusion models—systems trained on billions of parameters to predict text or generate images. Companies like OpenAI, Anthropic, Google, Meta, and dozens of others began training these systems and releasing them to the public. And unlike the platforms of the previous era, these AI systems had fundamentally different economics, requirements, and strategic logic.

An algorithmic feed, the core technology of platforms, can run indefinitely on marginal infrastructure costs. Once the algorithm is trained, serving additional users adds minimal cost. The value comes from adding users and extracting more data from their behavior. Open scaling—connecting more people, getting more data, training better algorithms—is the dominant strategy.

An AI system like GPT-4 or Claude, by contrast, is expensive to run. Each inference—each time a user asks a question—consumes compute resources and has a real marginal cost. Scaling to more users means scaling compute infrastructure. Training and operating these systems requires deep integration with proprietary data, specialized hardware, and domain-specific fine-tuning. The most valuable versions of these systems are not the largest or most general, but the most integrated with specific workflows and data. An AI system for medical diagnosis, trained on internal electronic health records and integrated with a hospital's IT infrastructure, is more valuable than a general-purpose system. A legal AI system trained on a specific law firm's case history and billing records is more useful than a general system. A research AI system with access to a university's computing clusters and proprietary datasets is more powerful than one running on public data.

These are not minor technical differences. They represent an inversion of the entire strategy that made platforms dominant. Where platforms benefited from open scaling and data extraction, AI systems benefit from closure and proprietary integration. Where platforms wanted users to switch between them as little as possible to maintain network effects, AI systems want to be deeply embedded in workflows—they want lock-in. Where platforms competed by building larger networks and extracting more behavioral data, AI systems compete by having access to better training data and deeper system integration.

The platform model could be distributed through an API. Google's search API, Facebook's social graph API, and Amazon's web services APIs allowed other companies to build on top of the platforms. The value of the platform came from its reach and ubiquity; sharing access through APIs didn't diminish that value—it extended it.

AI systems, by contrast, are not naturally distributed. The most valuable versions of these systems are proprietary. Sharing access through an API limits profitability when the real cost per inference is not zero. The value comes from proprietary training data, proprietary fine-tuning, proprietary integration. Companies racing to deploy AI systems are not thinking about how to distribute them through open APIs; they're thinking about how to lock customers into proprietary systems through deep integration and specialized features.

This shift is not something any individual company consciously decided would happen. It's the result of the technical and economic logic of AI systems being fundamentally different from the logic of platforms. But the effect is the same: a reversal, after two decades of expanding connection, toward fragmentation. A new era of walled gardens and specialized silos, where companies build AI systems tightly integrated with proprietary data and specific workflows, and where users must choose which ecosystem to participate in rather than benefiting from universal connection.

The Platform Bargain created the most connected era in human history. Two billion people could communicate across borders. Information could flow from anywhere to anywhere. A person in a small town could access the same content as someone in New York or Tokyo. Participation was democratized; anyone with a smartphone could build an audience or a business. But this universality came at a cost: centralization of power in the hands of a few companies, extraction of behavioral data, and algorithmic curation of information that often reinforced existing beliefs rather than broadening understanding.

The fragmentation that is beginning now is not a response to these problems. It is not being driven by a desire to democratize AI or return power to users. It is being driven by the economic logic of AI systems themselves, which are expensive, domain-specific, and profitable only when locked behind proprietary systems and integrated with specialized workflows. The next era will not be more decentralized than platforms; it will be less so. It will not offer universal access; it will offer specialized access to those who can pay. It will not be built on the exchange of behavioral data for free services; it will be built on subscription fees and proprietary integration.

To understand how this fragmentation is unfolding—and how complete it might become—requires examining not just the economic logic of AI systems, but how that logic is playing out in concrete domains where AI is being deployed at scale: medicine, creative work, science, cybersecurity, and labor. In each of these domains, the shift from platforms to AI systems is not just changing the technology; it is changing who has access to knowledge and capability, who profits from it, and who holds power over how it is used.