Andy Jassy, Satya Nadella, and Sundar Pichai now collect monopoly rent on every AI computation the economy performs.

It is true that the cloud businesses at Amazon, Microsoft, and Google posted numbers last quarter that investors could actually underwrite — AWS at roughly 37 percent revenue growth, Azure at 43 percent, Google Cloud at 82 percent — and that unlike the usual AI moonshots, cloud computing runs on a financial model analysts can parse. Buy or lease buildings and computing gear, rent it out via multi-year contracts, recoup the investment over a few years. Amazon disclosed a 39 percent operating margin on AWS, with average equipment payback under three years against customer contracts averaging five or more. The arithmetic works. Jassy suggested on the earnings call that AWS could eventually grow into a trillion-dollar annual revenue business, up from an estimated $170 billion this year. Amazon and Microsoft added roughly $950 billion in combined market value in a single week. Wall Street, after months of fretting about AI capex with no visible return, suddenly had a theory of value it could write down in a model.

The trouble is that the arithmetic describes a chokepoint, not a competitive market — and the numbers themselves demonstrate the distinction if you read past the investor-relations framing.

Cloud computing means that three firms have become the sole landlords of the infrastructure on which all computation now runs. When analysts note that “the ability to supply [computing] is more of a limitation than demand,” they are describing scarcity controlled by three suppliers — a market structure the Cicilline report from the House Antitrust Subcommittee documented across the dominant digital platforms in 2020, before the AI boom created a new class of captive customer. The $100-billion, ten-year contract AWS signed with Anthropic is presented as validation of demand. It is, structurally, a dependency relationship — a commitment so large and so long that Anthropic’s operational future is entangled with Amazon’s pricing decisions for the next decade. That is not a customer relationship. It is a lease. The AI labs are not voluntary customers in a competitive market. They are tenants in a landlord’s market, paying rents for compute they cannot get elsewhere, signed to five-year contracts whose terms the landlord wrote.

The engineering-substance discrimination here is worth being precise about, because the public discourse has the misleading habit of treating “the cloud” as a utility — a natural monopoly like a water main — rather than a concentration deliberately assembled over two decades through acquisition of potential competitors, suppression of interoperability, and a legal architecture of anti-circumvention that makes independent exit a legal hazard. The Computer Fraud and Abuse Act creates a minefield around any attempt to reverse-engineer interoperability with a cloud provider’s systems; the risk of prosecution is itself a deterrent that locks customers in regardless of price or performance. What Jassy describes as “predictable” infrastructure with “very compelling” returns on invested capital is, in plain terms, a toll booth on computation: three firms that collect rent on every calculation the economy performs, at prices they set, with exit costs they designed. The multi-year contracts, the proprietary APIs, the data-center footprints measured in gigawatts, the procurement relationships with foundries that take years to establish — these are not features of a competitive market. They are barriers to departure. And the “compelling return on invested capital” Jassy touts is monopoly rent, dressed in the language of capital efficiency. The 39 percent margin is not a return on capital. It is a return on the absence of exit.

There is a second axis of concentration that the “cloud = safe” narrative elides, and it is partly physical geography. The AI labs cannot move their workloads to smaller providers because smaller providers cannot secure the 500-megawatt data-center builds. Power availability and permitting cycles are real constraints, and the hyperscalers have locked them up — Microsoft’s twenty-year deal with Constellation for Three Mile Island Unit 1, Meta’s purchase of the entire 1.1-gigawatt Clinton Clean Energy Center in Illinois, Amazon adjacent to Talen’s Susquehanna nuclear plant. These are not contracts a new entrant can replicate. The natural monopoly is partly market structure and partly geology — and the two reinforce each other.

The four forces that historically kept infrastructure pricing honest — competition, regulation, self-help, and labour power — are each weakened at the cloud layer. There are not four competing cloud providers; there are three, and the barriers to entry include submarine-cable ownership and procurement pipelines that newer entrants cannot replicate at scale. Regulation exists but has not addressed cloud concentration as infrastructure: no proceeding analogous to the CRTC’s wholesale-access framework for telecommunications, no consultation on the cloud-equivalent of cable-landing-station ownership. Self-help is architecturally foreclosed: the APIs are proprietary, the contracts prohibit the adversarial interoperability that would let a competitor build alongside existing infrastructure, and the service agreements make reverse engineering a material breach. And the layoffs across the tech industry in 2023 and 2024 destroyed the labour leverage that once constrained enshittification from inside the firms — the engineer who could refuse an architecture choice because the market valued her scarce skills.

What Wall Street identified this week as the “key differentiator” between AI winners and losers is, more precisely, the infrastructure chokepoint that determines who collects rent regardless of whether any particular AI product delivers. The framing — cloud as a “tried-and-true business model” the hyperscalers could “fall back on” if the AI bubble bursts — gets the structure backward. The cloud is not a safety net beneath the AI bet. It is the extraction mechanism, and AI is the demand shock that justified the capital expenditure required to deepen the chokepoint. If Anthropic, OpenAI, and the other major buyers of AI computing falter, the contracts get reworked, the backlogs shrink, and the 39 percent margin becomes a 15 percent margin. But the lock-in persists, because the alternative is still to build your own data centres, your own networking, your own procurement pipeline — which is precisely the capital-intensive proposition that drove companies to the cloud in the first place. The extraction layer existed before AI and will exist after it. The rent survives any particular tenant.

The pattern is older than cloud computing. A company positions itself between a necessity and its users, raises the cost of departure until exit is economically irrational, then collects rent on the dependency it engineered. Gerdau of Porto Alegre did it to the steelworkers of Selkirk, Manitoba, when it acquired the town’s century-old mill in 1995 — kept the workers it could use, shed the ones it could not, changed the mill’s name without changing its logic. The mechanism was newer than the playbook. The hyperscalers are doing it to every entity that needs to compute. The apparatus is different — API, contractual lock-in, chip-procurement dependency, twenty-year power-purchase agreements — but the structure is the same: whoever controls the chokepoint sets the terms. Doctorow’s chokepoint-capitalism framework names the general pattern — intermediaries that squeeze both sides of a transaction, extracting value from the fact that neither can leave. The cloud is chokepoint capitalism applied to computation itself.

There is work being done on this. The FTC has opened inquiries into cloud-market competition. The EU’s Digital Markets Act includes interoperability provisions that, if enforced against cloud infrastructure, could begin to address the lock-in architecture. Submissions to these proceedings matter — not because they will produce a ruling next quarter, but because they establish the documentary record that subsequent enforcement will cite. The architecture is legible. The margins are in the quarterly filings. The switching costs are in the contract terms. The nuclear power-purchase agreements are in the SEC filings. The trillion-dollar number Jassy mentioned is not the forecast; it is the measure of what the extraction layer is worth when no one asks who owns the ladder. Deadlines are the only part of regulatory processes that the regulated actually respect, and submissions are the only part of regulatory records that subsequent governments have to read. The records are open. The work is to name the chokepoint before the rent becomes so embedded in the cost structure of computation that no intervention can extract it.