Monopolization and Competition

Imagine you founded a promising AI startup in 2024. You have a sharp team, a product people want, and early revenue. When it comes time to scale, you make a series of individually sensible decisions. You build on the most capable model's API. You host on the cloud provider offering the most generous credits. You take a strategic investment from a fund tied to one of the hyperscalers, because the money comes with priority access to scarce chips. Each choice is rational on its own. Together they add up to something you did not quite choose: your model supplier, your infrastructure provider, your largest investor, and — should they launch a competing feature — your fiercest rival are increasingly the same handful of companies. There is no clean exit, because there is nowhere independent to exit to.

This is not a thought experiment. It is the operating reality for a large share of the AI startups founded in the mid-2020s, and it sits at the center of the most consequential competition-policy debate of the decade. The question is not whether AI companies are innovating. They obviously are — the pace is dizzying. The question is who owns the ground the innovation stands on, and whether that ownership is being used to decide, quietly and in advance, who gets to compete at all.

What the Regulators Actually Did

In January 2024, the U.S. Federal Trade Commission opened a formal inquiry, sending compulsory orders to Microsoft, Google, Amazon, OpenAI, and Anthropic. It wanted the contracts — the investments, the cloud commitments, the board arrangements, the fine print. A year later, in January 2025, the FTC published its staff report on those partnerships. Its conclusion was couched in the careful conditional tense that regulators use when they have found something but have not yet decided to sue over it: the arrangements between cloud giants and AI developers raise "potential competition issues." The report identified concrete mechanisms — that the deals could shape access to scarce compute and engineering talent, raise the cost of switching between AI providers, and hand cloud companies unusual visibility into the operations of firms that might one day rival them.

That same summer, competition authorities on both sides of the Atlantic aligned. In July 2024, the FTC, the U.S. Department of Justice, the UK's Competition and Markets Authority, and the European Commission issued a rare joint statement naming three shared worries: concentrated control over the critical inputs of AI, the ability of entrenched incumbents to extend existing dominance into AI markets, and the competitive risks of the partnership structures themselves. Senators Elizabeth Warren and Ron Wyden had put the same concern in blunter terms in their investigation letters, warning that the deals could "discourage competition, circumvent our antitrust laws, and result in fewer choices and higher prices."

That is the state of play in one sentence: multiple governments have looked closely, agreed on where the pressure points are, and — so far — mostly stopped short of forcing anything to change.

How the Partnerships Are Built

To see why regulators are uneasy, you have to look at the architecture of the deals, which is more intricate than a headline investment figure suggests.

Microsoft's relationship with OpenAI is the archetype and, as of late 2025, the most revealing — because it was renegotiated. Microsoft had poured a reported $13 billion or more into OpenAI, in exchange for exclusive rights to host OpenAI's workloads on Azure, preferential access to its models, and the right to weave those models through Windows, Office, Bing, and GitHub. When OpenAI converted its capped-profit arm into a public-benefit corporation (the OpenAI Group) in October 2025, the terms were rewritten. Microsoft emerged holding roughly a 27% stake worth on the order of $135 billion — but its cloud exclusivity was gone. OpenAI became free to buy compute elsewhere, and promptly did, committing hundreds of billions of dollars to Oracle, tens of billions to Amazon's AWS, and striking chip and capacity deals with Nvidia, AMD, Broadcom, and CoreWeave. Microsoft kept intellectual-property rights to OpenAI's models and products through 2032, with any future declaration of "artificial general intelligence" — the trigger that would loosen those rights — now subject to verification by an independent expert panel rather than OpenAI's own say-so.

This restructuring matters for the antitrust question in a way that cuts in two directions at once, and honesty requires holding both. On one hand, it looks like competition working: exclusivity was pried loose, a second and third and fourth cloud entered the picture, and a genuinely independent AI lab now shops its enormous compute needs across suppliers. On the other, it is a reminder of how much leverage the incumbent retained — a decade of IP rights, a quarter of the equity, and a partner still deeply woven into its distribution.

The Anthropic side of the ledger shows the same pattern with two masters instead of one. Amazon has invested around $8 billion and become a primary infrastructure partner, building a dedicated cluster of its own Trainium chips for Anthropic's use. Google has invested billions more, taken a minority non-voting stake deliberately structured to stay under antitrust tripwires, and in late 2025 agreed to supply Anthropic with up to a million of its custom TPU accelerators — a commitment reportedly worth tens of billions of dollars. Anthropic, valued at roughly $183 billion in a September 2025 financing round, thus runs its models on hardware and clouds owned by the two companies that are simultaneously its largest backers and, through Amazon's and Google's own AI offerings, its competitors.

Deal	Approx. investment	Structure	Competitive entanglement
Microsoft–OpenAI	~$13B+ → ~27% stake (~$135B)	For-profit PBC; IP rights to 2032; cloud exclusivity ended 2025	Investor, longtime infra host, distributor via Office/Windows
Amazon–Anthropic	~$8B	Minority stake; dedicated Trainium compute	Investor, cloud/chip supplier, AI competitor (Bedrock, Nova)
Google–Anthropic	Billions; up to 1M TPUs	Non-voting minority stake; TPU supply	Investor, cloud/chip supplier, AI competitor (Gemini)

On paper, each of these is an ordinary venture arrangement: capital-hungry startups pairing with capital-rich platforms. What makes them different from an ordinary investment is the fusion of roles. When one company is your funder, your fuel supplier, and a rival selling into your market, the ordinary assumption behind competition — that firms sink or swim on their own merits — starts to wobble. This is why critics describe the structure as vertical integration in all but name. Nothing on the page says the AI labs must behave as satellites of their patrons. But the gravity is real.

The Mechanisms of Control

Why does that gravity persist even without an exclusivity clause to enforce it? Because the leverage in AI does not sit primarily at the model layer, where the headlines are. It sits underneath, in two strata that are far harder to reproduce than a clever model: chips and compute.

Start with the chips. Training and serving frontier models runs overwhelmingly on Nvidia's accelerators, which command something close to 90% of the market for AI data-center GPUs. This is not merely a lucrative position; it is a structural chokepoint whose effects propagate through the entire stack. Nvidia's advantage is not only its silicon but CUDA, the software layer that a generation of AI engineers learned on and that competitors' chips do not natively speak. When a single supplier controls the pickaxes in a gold rush this large, its allocation decisions become industrial policy by other means. Who gets chips, in what quantity, and at what price shapes which companies can even attempt to build at the frontier. The point sharpened in late 2025, when Nvidia began investing directly in the very customers who buy its hardware — most visibly a commitment reported at up to $100 billion tied to OpenAI's build-out. Money flowing from the chipmaker to a model lab that then spends it on the chipmaker's chips is either a virtuous flywheel or a circular arrangement that props up demand and blurs where independent judgment ends — and regulators, along with more than a few investors, have started asking which.

Above the chips sits compute. Training a frontier model means renting infrastructure at a scale that, in practice, only Amazon, Microsoft, and Google can supply, because no independent provider yet operates at comparable size. Challengers such as CoreWeave and Lambda have grown quickly, but they remain small against the hyperscalers and face the same brutal capital intensity that makes the market hard to enter in the first place. And the hyperscalers do not merely rent capacity; they already dominate the adjacent markets — search, cloud, advertising, e-commerce — that provide the proprietary data, the existing customer relationships, the bundling opportunities, and the deep balance sheets to absorb losses that would sink a smaller firm. Dominance in one layer subsidizes the fight for the next.

The Quiet Grip of Lock-In

Formal exclusivity, then, is almost beside the point. Lock-in does the work without it.

The most underappreciated mechanism is the compute credit. Cloud providers hand startups substantial credits and discounts to bring them onto a platform during the fragile early years. Those credits are not cash; they are pre-paid time on one company's infrastructure. A startup that has burned through a year of runway on Azure or AWS credits, built its stack around that provider's proprietary tools and optimized configurations, and trained its people on that environment, faces a migration that means rebuilding much of its foundation while keeping a live product running for paying customers. It is technically permitted. It is economically close to impossible. Most firms never try.

The same logic scaled all the way up explains why the Microsoft–OpenAI exclusivity, even once formally lifted, did not instantly scatter OpenAI's workloads to the winds. Years of engineering, tooling, and prepaid capacity are their own kind of tether. The FTC report flagged a further, subtler channel: talent. When the best engineers circulate within a small set of tightly integrated firms and their portfolio companies, expertise pools inside the incumbent networks rather than diffusing outward into the independent challengers that might otherwise erode their lead.

Does Open Source Break the Grip?

The most cited counterweight to all this is open-source AI, and the story has grown considerably more interesting than "Meta releases Llama."

The open-weight ecosystem is now genuinely global and genuinely capable. Meta's Llama models seeded it; OpenAI itself released open-weight models in 2025; and, most strikingly, Chinese labs vaulted to the frontier of the open field. DeepSeek's R1, released in January 2025, matched leading reasoning models at a fraction of the reported training cost and briefly rattled markets by suggesting the capital moat was thinner than assumed. Alibaba's Qwen family became one of the most downloaded and fine-tuned model lines in the world. Open weights let researchers, hobbyists, national governments, and small companies build systems that a closed ecosystem would never have permitted them — adapting models to local languages, sensitive data, and niche uses that no incumbent would prioritize.

Precise market-share figures are contested, and anyone quoting a single clean number is overstating what is known. What the available surveys suggest is a real and growing minority of deployments running on open-weight models, especially for fine-tuned, cost-sensitive, on-premises, or privacy-constrained workloads, while the highest-value frontier applications still lean heavily on proprietary APIs. The direction of travel — toward more open deployment — is clearer than any specific percentage.

But open source shifts the competitive terrain without resolving the underlying question of who holds the leverage. Meta can give Llama away precisely because its moat is social networking and advertising, not model weights; releasing the model wins developer goodwill, accelerates an ecosystem it benefits from, and buys regulatory credibility, all without surrendering anything that funds it. This is strategic generosity, not self-sacrifice. And it does nothing about the layers where leverage actually lives. Downloading a model is free. Fine-tuning it on proprietary data, serving it to millions of users, and keeping it running requires exactly the chips and compute controlled by the same firms under investigation. As one analysis put it, open-weight models "will not address the unregulated AI oligopoly at the hardware or cloud layers, and because model-layer enterprises are dependent on these lower layers, concentration means that oligopolists in these layers can leverage their power downstream." Making the software free does not change who owns the factory it runs in.

Whether open source is a durable structural equalizer or a temporary convenience for incumbents who feel no threat at the model layer is, honestly, not yet knowable. It is one of the genuine open questions in this field.

The Cases That Will Set the Rules

While regulators write reports, courts are producing something more binding: precedent. Three landmark cases against the incumbents will shape whether existing antitrust law reaches AI-era conduct — though each carries fact patterns specific enough that their transfer to AI is not guaranteed.

The Google search case is the most advanced and the most instructive. In 2024 a federal court found that Google had illegally monopolized general search. But in the remedies ruling that followed in September 2025, Judge Amit Mehta declined to order the structural breakup many expected — Google was not forced to divest Chrome. Instead the court barred certain exclusive default-placement contracts and required limited data-sharing with rivals, while allowing Google to keep making large payments to distribution partners. Tellingly, the judge cited the rise of AI assistants as a competitive force that reduced the need for harsher remedies — an argument that AI is loosening Google's grip even as Google races to dominate AI. Google is appealing. A separate case found Google liable in 2025 for monopolizing parts of the advertising-technology stack, with remedies still being fought over. The lesson so far echoes uncomfortably: liability is provable, but the remedy tends to arrive late and land soft.

Apple's App Store faces sustained legal pressure — from the long-running Epic Games litigation that forced it to permit external payment links, to the Justice Department's broader monopolization suit filed in 2024. The principle on trial is directly portable to AI: whether control of a platform, and the toll a gatekeeper can charge everyone who builds on it, crosses from advantage into abuse. Cloud infrastructure sits in a position uncannily like the App Store's, and a firm ruling against Apple's gatekeeping would strengthen the argument that infrastructure chokeholds have legal limits.

Amazon faces the FTC in a monopolization case in which the "Project Nessie" pricing algorithm — which the agency alleges was built to test how far Amazon could raise prices before rivals followed — features prominently. This is the case with the most direct bearing on AI, because it asks whether an algorithmic system can be an instrument of illegal market manipulation, and what legal standard would govern that finding. A ruling against Amazon would hand regulators a template for the coming wave of AI-driven pricing conduct.

Alongside the courts, legislatures are testing narrower tools. California's law restricting the use of shared "common pricing algorithms," which took effect in January 2026, targets exactly the scenario where competitors quietly converge on the same prices because they all lean on the same algorithm — collusion's result without collusion's smoking gun. The European Union, meanwhile, is weighing whether to treat dominant AI providers as "gatekeepers" under the Digital Markets Act, which could compel interoperability and forbid self-preferencing. Washington has moved the other way: the White House AI Action Plan of July 2025 leaned toward clearing regulatory obstacles to AI infrastructure rather than constraining it. The result is a widening transatlantic split, and with it the familiar risk of regulatory arbitrage, where firms route their structures toward the most permissive jurisdiction and pull the others toward leniency.

What Concentration Would Cost

Suppose the current trajectory holds and no intervention bites. The downstream effects are reasonably foreseeable, because we have watched concentrated platform markets before.

For startups, the viability calculus tightens. When your infrastructure, your best-in-class model, and a potential acquirer are the same three companies, the rational move for a founder is often to build something acquirable rather than something rivalrous — and the "kill zone" around incumbents, where venture capital avoids funding anything a hyperscaler might crush or copy, expands into AI. Innovation does not stop, but it narrows, channeled toward what complements the platforms rather than what challenges them. Pricing follows the usual arc of durable market power: introductory generosity while the field is contested, followed by the steady extraction that a locked-in customer base permits once switching is impractical. The compute credits that felt like gifts on the way in become leverage on the way out.

The second-order effects reach workers and consumers. A stack owned end to end by a few vertically integrated firms — controlling chips, compute, models, and the distribution channels into billions of devices — concentrates not just profit but bargaining power. Workers whose tools, and increasingly whose jobs, are mediated by these systems face employers with little competitive pressure to share the gains. Consumers inherit fewer genuine choices dressed as many, and prices set by firms that no longer need to fear a defector. This is the distribution question that runs through this whole book, arriving in a specific form: it is not mainly the technology that determines who benefits, but the market structure through which the technology reaches us.

What Real Competition Would Require

If open source is partial and enforcement is slow, restoring competitive conditions means structural intervention — feasible in engineering terms, hard in political ones — and it must be weighed against the real efficiencies that scale in AI genuinely provides. Training frontier models is expensive for physical reasons; some concentration reflects economies of scale that benefit everyone, not just abuse. The policy task is to preserve the former while curbing the latter, and the honest burden of proof should sit with the firms claiming that a given exclusivity, credit structure, or bundling arrangement produces efficiencies that could not be achieved through less restrictive means.

Four interventions do most of the work. Competitive access to compute is the most fundamental: public investment in shared compute infrastructure, on the model of roads, grids, and telecom networks, could give challengers a foundation they cannot currently rent at fair terms. Competition in chips addresses the deepest chokepoint: Nvidia's near-monopoly invites the kind of industrial policy the CHIPS Act aimed at fabrication, redirected toward AI accelerators and the open software standards that would let rival chips compete on merit rather than on CUDA compatibility. Interoperability and data portability attack lock-in directly: if workloads could move between clouds, models could interoperate through open standards, and customers could carry their data out the door, the switching costs that anchor incumbency would loosen — and because firms profit from those costs, none of this emerges without a mandate. Finally, transparency in the partnership contracts themselves is the precondition for the rest: regulators and the public cannot police exclusivity, pricing, or penalty terms they are not allowed to read.

The Uncomfortable Parallel

All of this rhymes with a moment the technology industry has lived through once already. In the late 1990s, Microsoft dominated personal computing, bundled Internet Explorer into Windows, and methodically starved Netscape. The Justice Department sued, won a finding that Microsoft had broken antitrust law — and then the case ended without the structural breakup originally sought. Microsoft was not split apart.

What happened next is the part worth sitting with. Microsoft kept its operating-system monopoly and lost the future anyway. Google took search. Apple took mobile. Amazon took cloud. The market moved to terrain where Microsoft's fortress gave it no advantage, and the dominance that looked permanent turned out to be a snapshot, not a sentence. The optimistic reading of today is that AI could stage the same escape: DeepSeek's cheap frontier model, the multi-cloud unwinding of the OpenAI deal, and the open-weight surge all hint that the moats are shallower than they look.

The pessimistic reading is that this time the incumbents are not one company caught flat-footed but the very firms that won the last three waves — and they have learned. They are running the recognizable playbook: extend platform dominance into the new market, lock partners in through infrastructure dependence, invest in promising rivals before they mature, and use scale to underprice anyone independent. The difference from the 1990s is that the chokepoints now sit in physical infrastructure — chips and data centers — that is far harder to route around than a browser.

Which reading wins turns on timing, and that is the sharpest uncertainty of all. Antitrust cases move in years; AI market structures set in months. The partnership scaffolding is already erected, the infrastructure already concentrated, the capital already committed. Enforcement that arrives after the arrangements have hardened tends, as the Microsoft case showed, to be enforcement that changes little. History is not destiny. But the window for shaping this market, rather than merely ratifying it, is narrowing with every quarter the current structure deepens.

Summary

Monopolization in AI is not a distant hazard; it is legible right now in the industry's architecture. A few firms — chiefly Microsoft, Google, Amazon, and, beneath them, Nvidia — control the chips, compute, and capital that determine who can compete. Their investments in the leading AI labs have created dependencies that function like vertical integration while remaining formally arm's-length.

Regulators have diagnosed the problem in unison. The FTC's 2025 report and the 2024 international joint statement named the same three chokepoints: concentrated control of critical inputs, incumbent entrenchment via adjacent dominance, and restrictive partnership structures. Landmark cases against Google, Apple, and Amazon are building the precedents that will decide whether existing law reaches AI conduct — though so far liability has come more readily than remedy.

The late-2025 restructuring of the Microsoft–OpenAI deal complicates any tidy verdict: exclusivity was loosened and rivals entered, which is what competition working looks like, even as the incumbent kept a quarter of the equity and a decade of rights. That ambiguity is the honest epistemic core of the chapter — we cannot yet cleanly distinguish partnerships that suppress competition from partnerships that enable it.

Open-source AI, now global and capable, genuinely widens access at the model layer but leaves the hardware and cloud bottlenecks untouched. Restoring competition would require public compute, chip-market competition policy, interoperability and data-portability mandates, and contract transparency — none of which exists at meaningful scale, and each of which must be weighed against the real economies of scale that make large-scale AI possible.

The Microsoft 1990s parallel offers both hope and warning: last time the market escaped a monopoly the law failed to break, but the chokepoints then were software, not silicon and data centers. Whether AI's bottlenecks prove as impermanent — and whether enforcement arrives before they harden — is the defining uncertainty.

Key Takeaways

Leverage lives below the model layer. The decisive chokepoints are Nvidia's roughly 90% grip on AI accelerators and the three hyperscalers' control of large-scale compute — not the models themselves, which is why open source alone cannot restore competition.
The partnerships function as vertical integration in all but name. When one firm is investor, infrastructure supplier, and competitor at once, genuine independence is hard to sustain — though the 2025 Microsoft–OpenAI restructuring shows the structure can loosen.
Lock-in works without exclusivity. Compute credits, migration costs, proprietary tooling, and concentrated talent networks tether firms to a provider even when they are formally free to leave.
Liability comes faster than remedy. The Google search ruling proved the monopoly but declined to break it up — the recurring pattern that lets market structure harden before enforcement bites.
Open source shifts the terrain without resolving it. A global, capable open-weight ecosystem (Llama, DeepSeek, Qwen) widens access but still runs on infrastructure owned by the firms under investigation.
The window is closing. Antitrust runs in years; AI markets set in months. Whether AI escapes concentration the way the market once outran Microsoft, or entrenches it in silicon and data centers, is the decade's open question.

Sources

Fresh web verification was unavailable for this revision; the following sources support the retained and updated claims and should be re-checked against the latest reporting.

Last updated: 2026-07-28

V2 (in progress) Previous: V1