You don't have an AI strategy. You have an exposure.

The most important fact about AI competition is also the least discussed. The market is not merely becoming concentrated: it is being built in a form that resembles the old trust structure, where control sits upstream in the infrastructure and downstream actors operate on terms set by a small number of gatekeepers.

That matters because ELECTE's remit is simple: who controls AI, who captures the value, and what that means for markets, competition, contracts, procurement, regulation, and labour.

Take that lens seriously and the current AI boom looks less like a broad market and more like a stacked system of dependency. Compute is concentrated. Foundation models are concentrated. Distribution is concentrated. Then subsidised access obscures the power structure long enough for dependency to harden into market architecture. I've called that dynamic The Subsidy IS the Strategy. In AI, cheap access often isn't generosity. It's a concentration mechanism.

The Return of the Trust

The old trust didn't just control a product category. It controlled the conditions under which the market could function. That is exactly why today's AI stack deserves to be read through the lens of monopolies and trusts.

What's being concentrated now is not only a software market. It is the infrastructure of prediction itself. The firms that sit closest to large-scale compute, proprietary data pipelines, foundation model development, and distribution channels don't merely sell services. They shape who else can enter, how cheaply they can enter, and whether they can keep any margin once they do.

That's a political economy problem before it's a technical one.

Control sits below the application layer

Many firms still talk about AI as though value will naturally accrue to whoever builds the most visible application. I don't buy it. In markets with concentrated infrastructure, application builders become tenants: they pay for access, depend on changing terms, and inherit strategic risk from suppliers they cannot discipline.

The central question is not who has the slickest interface. It is who owns the bottleneck.

That's why the language of monopolies and trusts is useful again. It directs attention away from feature demos and towards control over the system's essential layers.

Value capture follows infrastructure ownership

A competitive market distributes surplus across workers, suppliers, and customers. A concentrated stack pushes that surplus upwards.

The problem is not abstract. Research on AI capitalism describes a system marked by the commodification of data and concentration in AI talent and compute capacity, with the economic surplus enclosed by the entities that control those inputs rather than distributed across the labour market (Verdegem's analysis of AI capitalism).

Three points follow:

• Infrastructure ownership matters more than model theatre. Whoever controls compute and deployment conditions sets the practical limits of competition.
• Zero-price access can conceal dependency. A cheap API today can become a pricing or contractual lever tomorrow.
• Labour's position weakens when upstream control hardens. If the bottleneck sits in a few firms, the gains don't spread widely.

This isn't a story about innovation getting faster. It's a story about the terms on which intelligence is being industrialised.

The Original Antitrust Playbook

The trust problem isn't new. What is new is the substrate through which it operates.

In the late nineteenth century, industrial concentration produced business structures designed to suppress rivalry and coordinate power across firms. Trusts were, in substance, groups of companies acting together under legal agreements to reduce or threaten competition. Standard Oil became the defining example: by 1880 it controlled the refining of 90–95% of all oil produced in the United States, organised itself as a trust in 1882, and was dissolved by the Supreme Court in 1911 under the Sherman Act (Britannica).

Congress had responded with the Sherman Antitrust Act, approved on 2 July 1890 — the first federal law to outlaw monopolistic business practices. It declared any "contract, combination in the form of trust or otherwise, or conspiracy, in restraint of trade" illegal, authorised federal proceedings against trusts, imposed fines of $5,000 and up to a year in jail for individuals forming such combinations, and let private parties harmed by trusts sue for triple damages (U.S. National Archives on the Sherman Antitrust Act).

Senator John Sherman put the principle plainly:

"If we will not endure a king as a political power we should not endure a king over the production, transportation, and sale of any of the necessaries of life."

That line still holds. The object has changed from rail, oil, and steel to compute, data, and model access. The logic of concentrated control has not.

The trust-era economy produced two familiar concentration patterns: vertical integration, where a firm controls the chain from raw material to distribution, and horizontal integration, where a firm buys competitors at the same layer to fix prices or eliminate rivalry. The Clayton Antitrust Act of 1914 later strengthened the government's hand by closing loopholes the early courts had read into the Sherman framework.

The historical point isn't nostalgic. It is analytical. Antitrust began as a response to private power over essential economic systems — and that is the right lens for AI too. The difficulty is that the old playbook was written for visible combinations, sale prices, and industrial output. Today's forms of control are harder to count and easier to disguise.

Why Old Laws Falter in a Zero-Price World

The classic antitrust question was straightforward: did concentration let a firm raise prices, reduce output, or openly restrain trade in a defined market? That frame weakens when the service appears to be free and the bottleneck lies in data, defaults, distribution, and dependence.

Price is no longer the only relevant harm

A zero-price interface can still be economically extractive. Users may not pay cash, but they pay in data, attention, switching costs, and lock-in. A business customer may get discounted access at one layer while becoming strategically dependent at another. A narrow consumer-price test misses the underlying mechanism of control.

That matters in AI because the market is built around layers that reinforce each other. Data improves models. Models attract usage. Usage generates feedback. Distribution channels determine who gets discovered and who gets buried. The modern concern is not a visible cartel but a stack where one firm can shape access to infrastructure, terms of model use, distribution to users, the feedback loops that improve performance, and the commercial viability of everyone downstream. The consumer-welfare frame asks whether the short-run user price went up. It asks far less clearly whether the market has been made structurally dependent on a gatekeeper.

Practical rule: if a service is "free" but dependence grows every quarter, price is the wrong starting point.

Publisher economics offers a useful clue here. The conflict over AI crawling and content extraction has made it obvious that zero-price access can conceal one-sided value capture — I explored that dynamic in my piece on the pay-per-crawl shift in AI traffic monetisation. The same pattern applies more broadly: free access at the point of use can coincide with concentrated extraction at the point of infrastructure.

There's a further gap. Most monopoly analysis still centres on vertical and horizontal integration, yet the Yale Thurman Arnold Project highlights a neglected problem in common ownership — the same large investors holding significant stakes in competing firms, weakening competition without any traditional merger (Yale's Thurman Arnold Project on common ownership). That matters in AI because formal independence can coexist with aligned financial incentives. A market can look competitive in branding terms while remaining highly co-ordinated in structural terms.

The AI Stack as the New Monopoly

If you want to see where AI power sits, ignore the slogans and map the stack.

Start at the base. Per Synergy Research Group data for Q1 2026, AWS, Microsoft Azure, and Google Cloud together hold 63% of the worldwide cloud infrastructure market — 28%, 21%, and 14% respectively — with the rest of the field stuck at marginal shares, in a market that reached $129 billion in the quarter and is growing 35% year on year (Synergy data via Statista). Cloud and high-performance compute are the industrial base of AI. If access to that base is concentrated, every downstream market inherits the concentration.

In this context, the old monopoly vocabulary becomes precise again. The controlling firm doesn't need to own every application. It only needs to own the bottleneck that applications cannot route around. The distinction that matters is between a feature moat and an infrastructure moat — the latter is the one rivals cannot bypass.

Subsidy can be a weapon, not a gift

Cheap credits, below-cost access, bundled distribution, and generous onboarding terms are often framed as market expansion. Sometimes they are. But in concentrated infrastructure markets, subsidy can also be a device for absorbing demand before full pricing power arrives. A subsidised market can still be a monopolising market if the subsidy trains customers to rely on one stack, one route to users, and one supplier's economics.

In practice:

• At the compute layer, scale lowers the effective cost of experimentation for the incumbent.
• At the model layer, access terms can be relaxed early to attract developers, then tightened once dependence is established.
• At the distribution layer, the gatekeeper can privilege its own defaults and retain the user relationship.

I made a related point in my analysis of social networking aggregation: aggregation is not neutral when it determines who keeps the user, who sees the demand signal, and who can tax the interaction.

The AI capitalism literature gets the consequence exactly right. Concentration in data, talent, and compute doesn't just organise production. It organises surplus capture: the entity that controls the stack absorbs bargaining power from the firms above it and the labour below it.

When AI is sold as a universal productivity layer, ask a narrower question.

Who invoices for the bottleneck?

That question cuts through a lot of noise. It also explains why application markets can feel crowded while underlying power remains narrow.

Searching for a Modern Regulatory Response

Regulators have started to recognise the problem. The question is whether they are matching the structure of the market or merely reacting to its symptoms.

Europe is closer to the right frame

The European instinct has been to treat major digital actors as gatekeepers when their control over access creates market-wide influence — the logic behind the EU's Digital Markets Act. That framing is closer to present reality than a narrow obsession with short-run price effects, because it starts from intermediation power and dependency. The American approach has leaned more heavily on litigation, market-definition disputes, and proof burdens inherited from an earlier era. The live cases against major technology firms matter for exactly that reason: they test whether courts will treat distribution control, default power, and self-preferencing as structurally anti-competitive rather than merely aggressive business conduct.

The common policy line is that lawmakers need a whole new toolkit for platform and AI monopolies. I'm less convinced. The stronger argument comes from dormant authority already on the books. The Economic Security Project argues for unearthing the full arsenal of existing but dormant regulatory tools — Progressive-Era public-utility regulation that checks monopoly power through non-discrimination rules, and requirements for open and equal access and sufficient service — alongside the case law still on the books, rather than waiting for new legislation (Economic Security Project, Beyond Monopolies). That matters because AI concentration is moving faster than legislative cycles.

The hardest policy mistake is to confuse political delay with legal impossibility.

Where the bottleneck is infrastructural, conduct remedies alone can be too weak. A fine after the fact does little if dependency has already been built into procurement, contracts, and product design. A more realistic regulatory menu looks like this:

• Non-discrimination obligations for access to essential infrastructure.
• Structural separation where the operator of a bottleneck also competes on the layers that depend on it.
• Tighter merger scrutiny for acquisitions that extend control over adjacent AI layers.
• Attention to ownership structure, not only explicit combinations.

That is not anti-technology. It is a way of preventing infrastructure control from becoming a private constitution for the rest of the market.

Strategic Options for a Concentrated Market

If the AI stack is concentrating like a trust, then every actor needs a strategy for dependency.

For operators and mid-sized firms, the immediate risk is overcommitting to a single stack because the short-run pricing looks attractive. Procurement teams should treat model access, cloud access, and distribution access as governance questions, not only technical ones. If one supplier controls your margin, your roadmap, and your customer access, you don't have an AI strategy. You have an exposure.

A sensible operating posture has three parts:

• Diversify critical dependencies. Multi-model and multi-cloud planning is inconvenient, but concentration risk is rarely visible at the moment the contract is signed.
• Protect the feedback loop. Keep ownership and portability over the data, prompts, workflows, and evaluation layers that generate learning for your business.
• Separate experimentation from lock-in. Cheap access is useful for testing. It is dangerous when it imperceptibly becomes your permanent architecture.

For policymakers, the most serious bottleneck is compute. If that layer remains highly concentrated, the rest of the market keeps inheriting the same asymmetry. The Open Markets Institute makes the clearest structural proposal I've seen: recognise cloud computing as essential infrastructure, separate its ownership and control from the largest gatekeeper platforms, and regulate it as a public utility, so that AI's value does not accrue exclusively to private hyperscalers (Open Markets Institute, AI in the Public Interest). That sounds radical only if you still think cloud is just another private service line. In practice, it is closer to industrial infrastructure.

A related question is democratic control. Maximilian Kasy argues that fairness, privacy, and accountability cannot be settled without democratic control of the means of prediction — algorithm objectives, data, and computational infrastructure — through public debate and binding collective decision-making (INET Oxford working paper).

For corporate leaders, investors, and policymakers, the implication is the same. Don't read AI concentration as a temporary phase before competition naturally broadens. Read it as a structural tendency of the stack. I made a related argument in my piece on the cloud wars and AI's new infrastructure frontier: the key issue isn't which supplier wins the next cycle of enterprise spend. It's whether the market permits meaningful autonomy for everyone else.

Sources

• U.S. National Archives on the Sherman Antitrust Act
• Core Knowledge on monopolies, trusts, and pools
• Yale on modern antitrust enforcement
• SHS Conferences paper on AI market concentration
• PMC analysis of AI capitalism
• Economic Security Project on dormant antitrust tools
• Open Markets Institute on AI and public utility treatment for cloud
• Econstor paper on democratic control of the means of prediction

Fabio Lauria

CEO & Founder, ELECTE

Every week, we explore AI without the hype — using data, analysis and an independent perspective.