From toolmaker to toll collector
The most revealing fact about the AI boom is not how fast it has moved, but how quickly it has reassembled an old political economy. A technology sold as a universal productivity engine is turning into a market structure defined by choke points: model access, cloud distribution, proprietary data, app-store rules, and a small number of firms that can afford to control all of them at once.
That would be true even if the industry were calmer. It is not. OpenAI, Anthropic and Google now sit at the center of the market for general-purpose large language models, while Microsoft, Amazon and Google provide much of the underlying infrastructure, and Apple and Meta control the devices and social graphs through which AI is increasingly experienced. The result is a system in which the biggest companies do not merely compete to sell AI. They also compete to decide who else gets to sell it.
This is the part of the story that should worry policymakers most. For years, the standard antitrust concern in technology was that a platform would favor its own products over those of third parties. In AI, that dynamic is more severe because the platform is not just the marketplace; it is the model itself. If access to the model can be throttled, priced, degraded or revoked, then the platform’s leverage reaches deep into the products built on top of it.
That makes the AI sector feel less like the open internet of the 1990s and more like the era of vertically integrated utilities: a few suppliers with extraordinary control over an indispensable input. The difference is that this input is not electricity or steel. It is intelligence, or at least the commercially usable approximation of it.
The oligopoly at the core
What is striking about the market for general-purpose AI models is how quickly it has hardened into an oligopoly. OpenAI, Anthropic and Google dominate enterprise use, while the biggest cloud players supply the compute and the rest of the Big Tech ecosystem supplies distribution. In practice, this means that companies building AI applications are often dependent on the same firms they may someday need to compete against.
That dependency creates a simple but corrosive incentive. If a model provider sees a customer growing into a rival, it can raise costs, restrict access or cut off service altogether. If it sees a rival vulnerable to model dependence, it can use that dependence as a weapon. In older software markets, the concern was that incumbents copied features. In AI, they can contest the supply line itself.
Anthropic has already signaled that it is willing to police its model access with unusual firmness, including in cases where customers’ relationship to the company changed in ways it viewed as strategically sensitive. OpenAI and Google are not immune to similar temptations. Nor, for that matter, are Microsoft, Amazon and Meta, each of which has its own reasons to fold AI deeper into an existing business and therefore to blur the line between infrastructure provider, platform operator and competing application vendor.
The reason this matters is not merely that competition may become less fair. It is that AI markets may become less innovative. A startup will build differently if it knows the model can be withdrawn at any time. A large enterprise will adopt more slowly if the vendor supplying its core capabilities is also a competitor in its most valuable use case. The costs of dependence are not always visible in price sheets. They often show up in strategic caution, delayed launches and business models never attempted.
The cloud behind the curtain
Much of the public debate about AI has focused on model quality: benchmarks, hallucinations, reasoning, coding ability and speed. That is understandable, but incomplete. The real center of gravity is the stack beneath the model. Training frontier systems requires enormous capital expenditure in chips, data centers and energy. Running them at scale requires cloud infrastructure. Delivering them to consumers requires operating systems, browsers, app stores, search engines and social platforms.
Each layer reinforces the others. Google can integrate models into Search, Android and Workspace. Microsoft can bundle them into Office, Windows and Azure. Amazon can sell the compute that powers competitors while building its own AI products. Apple can decide which models are permitted, or visible, on the most lucrative consumer device on earth. Meta can use distribution through social feeds and messaging to normalize its own model layer, while also harvesting feedback from billions of interactions.
In a healthier market, those advantages would still matter, but they would not necessarily foreclose competition. In AI, they may. The product is general enough that a company can claim almost any use case. The distribution channels are sticky enough that users often accept defaults. And the economics are still opaque enough that firms with deep pockets can absorb losses for long enough to squeeze out challengers.
That dynamic is especially clear in consumer AI assistants. An assistant is not just another app; it is a layer that mediates search, commerce, communication and work. Whoever owns that layer can shape user behavior at scale. Whoever can insert the assistant into the operating system can do even more. Apple’s interest in making AI a native feature of the iPhone, Microsoft’s push to embed it in desktop software, Google’s attempt to protect search from being disintermediated, and Meta’s desire to keep people inside its own communication ecosystem all point to the same conclusion: AI is becoming a fight over where everyday attention begins and ends.
Competition by other means
Regulators have not been blind to this. The European Union has already taken a more assertive posture toward platform dominance, using the Digital Markets Act to push back on self-preferencing and gatekeeper behavior. In the United States, antitrust authorities have become more skeptical of Big Tech’s habit of combining platform control with adjacent product markets, though enforcement remains uneven and politically contested.
AI complicates matters because it sits at the intersection of several doctrines. It looks like a competition issue when a model provider cuts off access to a customer that may become a rival. It looks like a privacy issue when providers train on data scraped from the public web, licensed from publishers, or generated by users who have little practical control over downstream use. It looks like a consumer-protection issue when systems produce confident nonsense, hide their limitations or make it nearly impossible to understand how outputs were generated. And it looks like a national-security issue when frontier models are concentrated in a small number of American and Chinese firms with privileged access to chips, capital and talent.
That overlap gives regulators an opening, but also a trap. The opening is obvious: if traditional antitrust is too slow to address emergent bottlenecks, then ex ante rules on interoperability, portability, data access, and non-discrimination may be warranted. The trap is that regulators may focus on the most visible harms while missing the structural ones. A fine for privacy violations may not matter much if the underlying model market remains locked. A remedy against app-store favoritism may do little if the real power lies in the model provider’s ability to withhold access from its own developers.
The deeper question is whether regulators understand that AI is not one market but many stacked on top of each other. The temptation in Washington and Brussels is to treat models as if they were products. In reality, they are infrastructure, research pipelines, content engines, and strategic assets all at once. That makes them harder to govern, but also more dangerous to ignore.
The privacy bargain nobody explained
None of this would be politically sustainable without a second, quieter bargain: the exchange of convenience for data. AI systems improve when they are trained on more information, refined by more user interaction and embedded in more daily tasks. That creates a powerful appetite for personal, proprietary and behavioral data. It also creates ambiguity about consent.
Many users understand that a chatbot may remember a prompt or use conversations to improve service. Fewer understand how far that data can travel once it enters a model provider’s ecosystem, or how hard it may be to claw back. Enterprises worry about confidential material leaking into training pipelines. Consumers worry, more vaguely, that their digital lives are being vacuumed into systems they cannot inspect. Both worries are justified.
The privacy challenge in AI is not only collection. It is inference. Even if providers comply with formal deletion requests, models can retain statistical traces. Even if they strip obvious identifiers, prompts can reveal sensitive facts. Even if companies promise not to train on certain data, a system integrated across search, office software, messaging and cloud services can still build a remarkably detailed profile of a person or business. In a world where AI assistants are becoming default interfaces, the question is not whether data is collected. It is whether the user can still meaningfully understand and control what is being inferred from it.
That is why privacy rules designed for the web era look increasingly inadequate. Notice-and-consent regimes assume the user can choose among services and read the terms. AI systems are often bundled, opaque and ubiquitous. Data protection law may still help at the margins, but it will not on its own solve the deeper asymmetry between those who train the models and those whose lives become training material.
The old giants, newly menacing
Each of the Big Tech firms brings a different strategic logic to AI. OpenAI, though no longer a startup in any ordinary sense, remains the symbol of the frontier-model race and the company most associated with the consumer chatbot revolution. Anthropic has sold itself as the more restrained, safety-first competitor, but its willingness to enforce access boundaries shows that idealism is compatible with hard-nosed market power. Google is the incumbent with the most to lose if AI rewrites search, and the broadest set of assets with which to defend itself. Microsoft is the enterprise giant trying to turn AI into an indispensable layer of productivity software. Meta is using its distribution at scale to keep pace while betting that open-weight models and social products can undercut rivals. Apple is the cautious gatekeeper, eager to make AI feel native without ceding the device experience to anyone else.
That diversity of strategy is healthy in one sense. It suggests no single company has yet defined the industry’s future. But it also means competition is happening across several fronts at once, with no guarantee that the winners in one layer will not dominate the others. A model company can become a platform company. A platform company can become a model company. A cloud company can subsidize both. And each can shape the market by deciding which partners to empower and which to marginalize.
History suggests this is the moment when governments usually discover the architecture of their own weakness. By the time a market is obviously concentrated, the causal chain that produced the concentration is already hard to unwind. The firms have already hired the engineers, bought the chips, trained the models, secured the distribution and persuaded consumers that the service is indispensable. That is how bottlenecks become normal.
“The most valuable thing in AI may not be the model,” one industry veteran said recently, “but the right to decide who can use it.”
What a real response would look like
If governments want to preserve competition, they will need to be more imaginative than the standard antitrust toolkit allows. They should require clearer rules on model access for customers, especially when a provider also sells competing applications. They should push for portability across models and clouds so that a company can move its workloads without rebuilding its business from scratch. They should scrutinize exclusive deals that bind model providers to cloud incumbents or device makers. And they should insist that privacy protections apply not only to the data visibly collected, but also to the inferences made from it.
In some cases, structural remedies may be necessary. The most obvious is separating model providers from downstream applications where the conflict of interest is severe. Another is treating large AI systems more like essential infrastructure, subject to non-discrimination obligations. A third is demanding transparency around training data, fine-tuning data and access restrictions, so that customers understand what they are buying and what can be taken away.
None of these moves will be easy. Each would be opposed by firms that argue innovation requires flexibility and that safety sometimes requires control. They are not entirely wrong. AI systems are still fragile, prone to abuse and expensive to run. Providers do need some discretion to prevent misuse, protect intellectual property and maintain security. But discretion is not the same as arbitrary power. A market built on dependence should not be allowed to normalize the threat of sudden exclusion.
The broader political challenge is that AI is arriving at a moment of exhaustion. Voters are skeptical of Big Tech, but they also want useful tools. Governments want growth, but they fear concentration. Companies want speed, but they fear liability. That makes it tempting to postpone the hardest questions until the market matures. Yet the market will not mature in a neutral direction. It will mature toward whatever structure is being built now.
If the first decade of modern internet regulation was about content, the second was about platforms, and the third may be about models. The underlying issue is the same in each case: when a small number of firms control the infrastructure through which everyone else must pass, private incentives start to look uncomfortably like public power. AI is merely giving that old lesson a more advanced vocabulary.
The industry likes to speak in the language of abundance, and for good reason. The systems are astonishing. They write, summarize, code, search, converse and predict with increasing fluency. But abundance in capability does not automatically produce abundance in power. Sometimes it does the opposite. When intelligence becomes a product, the firms that supply it may end up knowing more, controlling more and letting others do less.
That is the real story of AI in 2026. Not that machines are taking over, but that the companies building them are discovering how to turn possibility itself into leverage. If regulators fail to see that clearly, the next great technological revolution may also become the most concentrated one yet.
